{"id":108,"date":"2020-09-03T16:39:48","date_gmt":"2020-09-03T13:39:48","guid":{"rendered":"http:\/\/acikkitap.com.tr\/oaek\/chapter\/bolum-19-granuler-buyuk-olcekli-eylem-verileri-kullanarak-ogrenci-davranisinin-kestirimci-modellemesi\/"},"modified":"2020-09-03T16:39:48","modified_gmt":"2020-09-03T13:39:48","slug":"bolum-19-granuler-buyuk-olcekli-eylem-verileri-kullanarak-ogrenci-davranisinin-kestirimci-modellemesi","status":"publish","type":"chapter","link":"https:\/\/acikkitap.com.tr\/oaek\/chapter\/bolum-19-granuler-buyuk-olcekli-eylem-verileri-kullanarak-ogrenci-davranisinin-kestirimci-modellemesi\/","title":{"raw":"B\u00f6l\u00fcm 19 Gran\u00fcler B\u00fcy\u00fck \u00d6l\u00e7ekli Eylem Verileri Kullanarak \u00d6\u011frenci Davran\u0131\u015f\u0131n\u0131n Kestirimci Modellemesi","rendered":"B\u00f6l\u00fcm 19 Gran\u00fcler B\u00fcy\u00fck \u00d6l\u00e7ekli Eylem Verileri Kullanarak \u00d6\u011frenci Davran\u0131\u015f\u0131n\u0131n Kestirimci Modellemesi"},"content":{"raw":"\n

Steven Tang, Joshua C. Peterson ve Zachary A. Pardos<\/span><\/span><\/p>\n

E\u011fitim Bilimleri Enstit\u00fcs\u00fc, UC Berkeley, ABD<\/span><\/span><\/p>\n

DOI: 10.18608\/hla17.019<\/span><\/span><\/p>\n\n

\u00d6Z<\/h2>\n

Kitlesel a\u00e7\u0131k \u00e7evrimi\u00e7i dersler (KA\u00c7D), \u00f6\u011frenenlerin \u00f6\u011frenme materyalleriyle etkile\u015fime girerken ve kavrama etkinliklerini tamamlarken g\u00f6sterdikleri eylemlerin kapsaml\u0131 bir kayd\u0131n\u0131 olu\u015fturur. Bu y\u00fcksek hacimli s\u0131ral\u0131 veri ve se\u00e7im ile \u00f6\u011frenci davran\u0131\u015f\u0131n\u0131 modelleme potansiyeli gelmektedir. \u00d6\u011frenme ortamlar\u0131ndan kaydedilenler gibi uzun vadeli, s\u0131ral\u0131 verilere bakmak i\u00e7in \u00e7e\u015fitli y\u00f6ntemler vard\u0131r. Dil modellemesi alan\u0131nda, geleneksel n-gram teknikleri ve modern tekrarlayan sinir a\u011flar\u0131 (TSA) yakla\u015f\u0131mlar\u0131, dilde yap\u0131y\u0131 algoritmik olarak bulmak ve \u00f6nceki kelimeleri girdi olarak verilen c\u00fcmle veya paragrafta bir sonraki s\u00f6zc\u00fc\u011f\u00fc tahmin etmek i\u00e7in uygulanmaktad\u0131r. Bu b\u00f6l\u00fcmde biz bu \u00e7al\u0131\u015fmaya, bir KA\u00c7D'deki kaynak g\u00f6r\u00fcn\u00fcmleri ve etkile\u015fimlerinin \u00f6\u011frenci dizilimlerini girdiler olarak ele alarak ve \u00f6\u011frencilerin bir sonraki etkile\u015fimini \u00e7\u0131kt\u0131lar olarak tahmin ederek bir benzetim yap\u0131yoruz. <\/span><\/span><\/p>\n

Bizim yakla\u015f\u0131m\u0131m\u0131z, KA\u00c7D'de kaynaklar\u0131n temsilini belirgin bir \u00f6zellik m\u00fchendisli\u011fi1<\/a><\/sup> gerektirmeden \u00f6\u011frenir. Bu model potansiyel olarak, bir \u00f6\u011frencinin ba\u015far\u0131 elde etmek i\u00e7in bir sonraki ad\u0131mda yapmas\u0131 gereken eylemlere dair \u00f6neriler \u00fcretmek i\u00e7in kullan\u0131labilir. Ek olarak, b\u00f6yle bir model otomatik olarak performans ve duyu\u015f hakk\u0131nda \u00e7\u0131kar\u0131m sa\u011flayan bir \u00f6\u011frenci davran\u0131\u015fsal durumu olu\u015fturur. \u00c7al\u0131\u015fmam\u0131zda kullan\u0131lan KA\u00c7D\u2019nin 3.500\u2019den fazla e\u015fsiz kayna\u011f\u0131 oldu\u011fu g\u00f6z \u00f6n\u00fcne al\u0131nd\u0131\u011f\u0131nda, bir \u00f6\u011frencinin etkile\u015fime girece\u011fi bir sonraki kesin kayna\u011f\u0131 tahmin etmek zor bir s\u0131n\u0131fland\u0131rma problemi gibi g\u00f6r\u00fcnebilir. Ders program\u0131n\u0131n (dersin yap\u0131s\u0131n\u0131n) bu \u00f6ng\u00f6r\u00fcy\u00fc yapmada ortalama %23 do\u011fruluk sa\u011flad\u0131\u011f\u0131n\u0131, ard\u0131ndan %70.4 ile n-gram y\u00f6ntemini ve %72.2 ile TSA bazl\u0131 y\u00f6ntemleri takip etti\u011fini ke\u015ffettik. Bu ara\u015ft\u0131rma, \u00f6zellik m\u00fchendisli\u011fi gerektirmeyen teknikler kullanarak ince taneli zaman serisi \u00f6\u011frenci verilerinin davran\u0131\u015f modellemesi i\u00e7in zemin haz\u0131rlamaktad\u0131r.<\/span><\/span><\/p>\n

Anahtar Kelimeler<\/span>: Davran\u0131\u015f modellemesi, dizilim tahmini, KA\u00c7D'ler, TSA'lar<\/span><\/span><\/p>\n

G\u00fcn\u00fcm\u00fcz\u00fcn dijital d\u00fcnyas\u0131, b\u00fcy\u00fck kullan\u0131c\u0131 eylemlerinin g\u00fcnl\u00fck kay\u0131tlar\u0131na dayanan ki\u015fiselle\u015ftirme ile i\u015faretlenmi\u015ftir. E\u011fitim alan\u0131nda, kullan\u0131c\u0131n\u0131n (genellikle gizli) \u00f6zelliklerine g\u00f6re \u00f6\u011frenme \u00f6nerilerini ve sonu\u00e7lar\u0131n\u0131 bireysel kullan\u0131c\u0131lara uyarlayabilen ki\u015fiselle\u015ftirilmi\u015f ve otomatik \u00f6\u011freticilere y\u00f6nelik ara\u015ft\u0131rmalar devam etmektedir. Son y\u0131llarda, kitlesel a\u00e7\u0131k \u00e7evrimi\u00e7i dersler (KA\u00c7D'ler) gibi y\u00fcksek\u00f6\u011frenim \u00e7evrimi\u00e7i \u00f6\u011frenme ortamlar\u0131 \u00f6\u011frenci taraf\u0131ndan olu\u015fturulan y\u00fcksek miktarda \u00f6\u011frenme eylemlerini bir araya getirmi\u015ftir. Bu b\u00f6l\u00fcmde, \u00f6\u011frenmeyi istendi\u011fi kadar eri\u015filebilir, sa\u011flam ve verimli k\u0131lmak i\u00e7in \u00f6\u011frenme yollar\u0131n\u0131 ki\u015fiselle\u015ftirme becerisine y\u00f6nelik, \u00f6\u011frenci taraf\u0131ndan olu\u015fturulan b\u00fcy\u00fck veri kaynaklar\u0131n\u0131 kullanmay\u0131 ama\u00e7layan ve giderek b\u00fcy\u00fcyen ara\u015ft\u0131rma alan\u0131na katk\u0131da bulunmaya \u00e7al\u0131\u015f\u0131yoruz. Bunu yapmak i\u00e7in, \u00f6ncelikle performans de\u011ferlendirme ve tahmin ile ilgili ara\u015ft\u0131rma hedeflerinden farkl\u0131 olarak, \u00f6\u011frencinin davran\u0131\u015fsal durumunun modellenmesine odaklanan bir ara\u015ft\u0131rma dizisi g\u00f6steriyoruz. Bir KA\u00c7D\u2019deki \u00f6\u011frencilerin ders videolar\u0131n\u0131 izlemek ya da forum yaz\u0131lar\u0131na cevap vermek ve bir sonraki eylemlerini tahmin etmek gibi t\u00fcm eylemlerini g\u00f6z \u00f6n\u00fcnde bulundurmak istiyoruz. B\u00f6yle bir yakla\u015f\u0131m, KA\u00c7D'larda toplanan ayr\u0131nt\u0131l\u0131, de\u011ferlendirme d\u0131\u015f\u0131 verileri kullan\u0131r ve seyir rehberli\u011fi arayan \u00f6\u011frenciler i\u00e7in bir tavsiye kayna\u011f\u0131 olarak hizmet etme potansiyeline sahiptir.<\/span><\/p>\n

KA\u00c7D'lere kat\u0131lan on binlerce \u00f6\u011frencinin t\u0131klama ak\u0131\u015f verilerini kullanarak, dersin sonunda ba\u015far\u0131l\u0131 olanlar\u0131n davran\u0131\u015flar\u0131n\u0131 modelleyerek, KA\u00c7D'ler aras\u0131nda gezinen \u00f6\u011frenciler aras\u0131nda genelle\u015ftirilebilir eylem \u00f6r\u00fcnt\u00fclerinin, ortaya \u00e7\u0131k\u0131p \u00e7\u0131kamayaca\u011f\u0131n\u0131 soruyoruz. Ba\u015far\u0131l\u0131 \u00f6\u011frencilerin KA\u00c7D'ler \u00fczerindeki trendlerini yakalamak, otomatik \u00f6neri sistemlerinin geli\u015ftirilmesini sa\u011flayabilir, b\u00f6ylece zorlanan \u00f6\u011frencilere ba\u015far\u0131l\u0131 olmak i\u00e7in harcad\u0131klar\u0131 zaman\u0131 optimize etmek i\u00e7in anlaml\u0131 ve etkili \u00f6neriler verilebilir. Bu g\u00f6rev i\u00e7in \u00fcretici s\u0131ral\u0131 modellerden yararlan\u0131r\u0131z. \u00dcretici dizilimli modeller girdi olarak bir olaylar diziliminde yer alabilir ve daha sonra ger\u00e7ekle\u015fecek olay \u00fczerinde olas\u0131l\u0131k da\u011f\u0131l\u0131m\u0131 olu\u015fturabilir. Bu \u00e7al\u0131\u015fmada di\u011fer \u00fcretici ve s\u0131ral\u0131 g\u00f6revlere uyguland\u0131\u011f\u0131nda geleneksel olarak ba\u015far\u0131l\u0131 olan \u00f6zellikle n-gram ve tekrarlayan sinir a\u011f\u0131 (TSA) modelleri olmak \u00fczere \u00fcretici s\u0131ral\u0131 modellerden iki t\u00fcr kullan\u0131lm\u0131\u015ft\u0131r.<\/span><\/p>\n

Bu b\u00f6l\u00fcm \u00f6zellikle, \u00f6\u011frencinin bir KA\u00c7D\u2019de ger\u00e7ekle\u015ftirdi\u011fi \u00f6nceki eylemler ba\u011flam\u0131nda verilen bir sonraki eylemi ne kadar iyi tahmin edebilece\u011fini analiz eder. Bu t\u00fcr bir analizin amac\u0131 nihayetinde, otomatik bir dan\u0131\u015fman\u0131n, \u00f6\u011frenciyi sonradan hangi eylemde bulunabilece\u011fi konusunda anlaml\u0131 bir rehberlik sa\u011flamak i\u00e7in modeli sorgulayabilece\u011fi bir sistem olu\u015fturmak olacakt\u0131r. Bir \u00e7ok durumda bir sonraki eylem,ders taraf\u0131ndan \u00f6ng\u00f6r\u00fclen bir sonraki kaynak olabilir ancak di\u011fer durumlarda, \u00f6\u011frencinin bilmedi\u011fi bir ders kitab\u0131n\u0131n bir k\u00f6\u015fesinde g\u00f6m\u00fcl\u00fc olan bir \u00f6nceki dersten veya zenginle\u015ftirme materyalinden bir kayna\u011fa ba\u015fvurmak bir tavsiye olabilir. E\u011fitti\u011fimiz bu bu modeller, \u00fcretici2<\/a><\/sup> olarak bilinirler, \u00e7\u00fcnk\u00fc \u00f6\u011frencinin daha \u00f6nce hangi eylemleri ger\u00e7ekle\u015ftirdi\u011fine ili\u015fkin \u00f6nceki bir ba\u011flam dikkate al\u0131nd\u0131\u011f\u0131nda hangi eylemin gelebilece\u011fini olu\u015fturmak i\u00e7in kullan\u0131labilirler. Eylemler, ders videosu a\u00e7ma, s\u0131nav sorusu cevaplama ya da bir forum g\u00f6nderisinde gezinme ve cevaplama gibi \u015feyleri i\u00e7erebilir. Bu ara\u015ft\u0131rma, KA\u00c7D'lerde potansiyel uygulamalarla ki\u015fiselle\u015ftirilmi\u015f dan\u0131\u015fmanlar olu\u015fturmaya y\u00f6nelik s\u0131ral\u0131, \u00fcretici modellerin birbirini takip eden verilerle di\u011fer e\u011fitim ba\u011flamlar\u0131na uygulanmas\u0131 i\u00e7in bir temel olarak hizmet vermektedir.<\/span><\/p>\n\n

\u0130LG\u0130L\u0130 \u00c7ALI\u015eMALAR<\/h2>\n

\u0130ngilizce dili s\u00f6z konusu oldu\u011funda, \u00fcretici modeller \u00f6rnek metinler olu\u015fturmak ve bu dilin nas\u0131l yap\u0131land\u0131r\u0131ld\u0131\u011f\u0131na dair \u00f6rnek metnin akla yatk\u0131nl\u0131\u011f\u0131n\u0131 de\u011ferlendirmek i\u00e7in kullan\u0131l\u0131r. Do\u011fal dil i\u015flemede kullan\u0131lan basit fakat g\u00fc\u00e7l\u00fc bir model (DD\u0130), olas\u0131l\u0131k da\u011f\u0131l\u0131m\u0131n\u0131n e\u011fitim k\u00fcmesinde her olas\u0131 n terim diziliminin \u00fczerine \u00f6\u011frenildi\u011fi bir n-gram modeldir (Brown, Desouza, Mercer, Pietra ve Lai, 1992). Son zamanlarda, tekrarlayan sinir a\u011flar\u0131 (TSA'lar), daha \u00f6nce g\u00f6r\u00fclen kelimelerin y\u00fcksek boyutlu s\u00fcrekli gizil bir duruma getirildi\u011fi bir sonraki kelime tahminini (Mikolov, Karafiat, Burget, Cernocky ve Khudanpur, 2010) ger\u00e7ekle\u015ftirmek i\u00e7in kullan\u0131lm\u0131\u015ft\u0131r. Bu gizli durum, daha \u00f6nce ba\u011flamda g\u00f6r\u00fclen kelimelerin hepsinin \u00f6zl\u00fc bir say\u0131sal temsilidir. Model daha sonra hangi kelimelerin geleceklerini tahmin etmek i\u00e7in bu g\u00f6sterimi kullanabilir. Bu \u00fcretici modellerin her ikisi de c\u00fcmleleri tamamlamak amac\u0131yla aday c\u00fcmleler ve kelimeler \u00fcretmek i\u00e7in kullan\u0131labilir. Bu \u00e7al\u0131\u015fmada, kelime ve c\u00fcmle dizilimlerinin akla yak\u0131nl\u0131k durumunu \u00f6\u011frenmek yerine, \u00fcretici modeller KA\u00c7D ba\u011flamlar\u0131nda \u00f6\u011frencilerin \u00fcstlendikleri eylem dizilimlerinin uygunlu\u011funu \u00f6\u011freneceklerdir. Daha sonra, bu t\u00fcr \u00fcretici modeller, \u00f6\u011frencinin daha sonra yapmas\u0131 gerekenler i\u00e7in tavsiyeler \u00fcretmek i\u00e7in kullan\u0131labilir.<\/span><\/p>\n

\u00d6\u011frenme analiti\u011fi toplulu\u011funda, \u00e7o\u011fu zaman KA\u00c7D ba\u011flamlar\u0131nda, \u00f6\u011frenciler taraf\u0131ndan \u00fcretilen verilerin analiz edildi\u011fi ilgili bir \u00e7al\u0131\u015fma vard\u0131r. Analitikler bir\u00e7ok farkl\u0131 \u00f6\u011frenci taraf\u0131ndan olu\u015fturulan veri t\u00fcr\u00fcyle ger\u00e7ekle\u015ftirilir ve pek \u00e7ok farkl\u0131 t\u00fcrde tahmin g\u00f6revi vard\u0131r. Crossley, Paquette, Dascalu, McNamara ve Baker (2016), bu durum i\u00e7in KA\u00c7D\u2019den al\u0131nan ham g\u00fcnl\u00fck kay\u0131tlar\u0131n\u0131n manuel \u00f6zellik m\u00fchendisli\u011fi s\u00fcreci ile \u00f6zetlendi\u011fi bir paradigma \u00f6rne\u011fi sunmaktad\u0131r. Bizim yakla\u015f\u0131m\u0131m\u0131zda, \u00f6zellik g\u00f6sterimleri do\u011frudan ham zaman serisi verilerinden \u00f6\u011frenilir. Bu yakla\u015f\u0131m, \u00f6zelliklerin geli\u015ftirilmesi i\u00e7in konu uzmanl\u0131\u011f\u0131 gerektirmez ve KA\u00c7D t\u0131klama ak\u0131\u015f\u0131ndaki ham bilgileri kullanmak i\u00e7in potansiyel olarak daha az kayb\u0131 olan bir yakla\u015f\u0131md\u0131r. Pardos ve Xu (2016), \u00f6nceki bilgilerin, KA\u00c7D kaynak kullan\u0131m\u0131 ile bilgi edinimi aras\u0131ndaki ili\u015fkinin geli\u015ftirilmesine yard\u0131mc\u0131 olmakta zorland\u0131\u011f\u0131n\u0131 belirledi. Bu \u00e7al\u0131\u015fmada, \u00f6\u011frencinin kendi kendini se\u00e7me durumu bir parazit ses ve ak\u0131l kar\u0131\u015ft\u0131r\u0131c\u0131l\u0131k kayna\u011f\u0131d\u0131r. Buna kar\u015f\u0131l\u0131k, \u00f6\u011frenen se\u00e7imi davran\u0131\u015fsal modellemede bir i\u015faret haline gelir. Reddy, Labutov ve Joachims (2016) 'da, \u00e7evrimi\u00e7i bir ders sistemindeki \u00f6\u011frenci \u00f6\u011frenmesinin bir\u00e7ok y\u00f6n\u00fc, g\u00f6mme3<\/a><\/sup> yoluyla birlikte \u00f6zetlenmi\u015ftir. Bu g\u00f6mme i\u015flemi \u00f6devleri, \u00f6\u011frenci becerisini ve ders etkinli\u011fini d\u00fc\u015f\u00fck boyutlu bir uzaya e\u015fler. B\u00f6yle bir s\u00fcre\u00e7, modelin mevcut \u00f6\u011frenci yetene\u011fi tahminine dayanarak ders verme ve \u00f6dev verme yollar\u0131n\u0131n \u00f6nerilmesini sa\u011flar. Bu b\u00f6l\u00fcmdeki \u00e7al\u0131\u015fma ayn\u0131 zamanda \u00f6\u011frenciler i\u00e7in \u00f6\u011frenme yollar\u0131 \u00f6nermeyi ama\u00e7lamaktad\u0131r ancak forum sonras\u0131 eri\u015fimler ve ders video g\u00f6r\u00fcnt\u00fclemeleri gibi ek \u00f6\u011frenci davran\u0131\u015flar\u0131n\u0131n da modele d\u00e2hil edilmesi ile farkl\u0131la\u015f\u0131rlar. Ek olarak, farkl\u0131 \u00fcretici modeller kullan\u0131l\u0131r.<\/span><\/p>\n

Bu b\u00f6l\u00fcmde, yaln\u0131zca KA\u00c7D'lerden gelen kay\u0131t g\u00fcnl\u00fc\u011f\u00fc verileriyle \u00e7al\u0131\u015f\u0131yoruz. Bu kullan\u0131c\u0131 t\u0131klama ak\u0131\u015f\u0131 bir\u00e7ok etkile\u015fim alan\u0131n\u0131n \u00fcst\u00fcnden ge\u00e7erken, davran\u0131\u015f ara\u015ft\u0131rmas\u0131 \u00f6rnekleri bu etkile\u015fim dizilimlerine kat\u0131lan kaynaklar\u0131n i\u00e7eri\u011fini analiz etmi\u015ftir. Bu \u00f6rnekler, videonun etkile\u015fim d\u00fczeyini (Sharma, Biswas, Gandhi, Patil ve Deshmukh, 2016) karakterize etmek i\u00e7in KA\u00c7D video karelerinin analiz edilmesini, forum yaz\u0131lar\u0131n\u0131n i\u00e7eri\u011fini (Wen, Yang ve Rose, 2014; Reich, Stewart, Mavon ve Tingley, 2016) ve forumlardaki etkile\u015fimlerden kaynaklanan ve buna mahsus sosyal a\u011flar\u0131n analizini i\u00e7erir (Oleksandra ve Shane, 2016). T\u00fcm olas\u0131 \u00f6\u011frenci etkinlikleri kategorilerine bu i\u00e7erik odakl\u0131 yakla\u015f\u0131mlara k\u0131yasla daha soyut bir d\u00fczeyde bak\u0131yoruz.<\/span><\/p>\n

\u00d6\u011frenme analiti\u011fi ve EVM'deki bili\u015f a\u00e7\u0131s\u0131ndan, Bayesci bilgi takibi (BBT; Corbett ve Anderson, 1994) y\u00fczeysel ders yap\u0131s\u0131n\u0131 bir bilgi bile\u015fenleri kayna\u011f\u0131 olarak kullanarak modelin bir KA\u00c7D'ye iyile\u015ftirilerek uyarlanmas\u0131(Pardos, Bergner, Seaton ve Pritchard, 2013) gibi modeller arac\u0131l\u0131\u011f\u0131yla \u00f6\u011frencilerin mahrem bilgilerinin de\u011ferlendirilmesi i\u00e7in bir\u00e7ok \u00e7al\u0131\u015fma yap\u0131lm\u0131\u015ft\u0131r. Bu modelleme t\u00fcr\u00fc, \u00f6\u011frencilerin davran\u0131\u015flar\u0131n\u0131 \u00f6\u011frencilerin gizli bilgisini modellemek i\u00e7in \u00f6\u011frenme f\u0131rsatlar\u0131 olarak g\u00f6r\u00fcr. \u00c7al\u0131\u015fma ile ilgili olmas\u0131na ra\u011fmen, \u00f6\u011frenci bilgisi bu b\u00f6l\u00fcmde a\u00e7\u0131k\u00e7a modellenmemi\u015ftir. Bunun yerine, modellerimiz, \u00f6\u011frencinin davran\u0131\u015f verileri olan bu performans verilerinin tamamlay\u0131c\u0131s\u0131n\u0131 tahmin etmeye odaklan\u0131r.<\/span><\/p>\n

Derin bilgi takibi (DBT; Piech vd., 2015), tekrarlayan sinir a\u011flar\u0131n\u0131n, \u00f6\u011frencilerin \u00e7evrimi\u00e7i \u00f6\u011frenme ortamlar\u0131nda gezinirken daha \u00f6nce g\u00f6r\u00fclen de\u011ferlendirme sonu\u00e7lar\u0131na dayanarak s\u00fcrekli gizli bir temsilini olu\u015fturmak i\u00e7in kullan\u0131r. Bu \u00e7al\u0131\u015fmada, tekrarlayan sinir a\u011flar\u0131 karma\u015f\u0131k bir gizli durumu izleyerek bir \u00f6\u011frencinin \u00f6nceki de\u011ferlendirme sonu\u00e7lar\u0131n\u0131 \u00f6zetlemektedir. Bu \u00e7al\u0131\u015fma, y\u00fczeysel BBT yakla\u015f\u0131m\u0131na k\u0131yasla \u00f6\u011frenci bilgilerini temsil etmek i\u00e7in derinlemesine bir \u00f6\u011frenme yakla\u015f\u0131m\u0131n\u0131n kullan\u0131labilece\u011fini g\u00f6stermektedir. Bununla birlikte, bu sonu\u00e7lar\u0131n h\u00e2lihaz\u0131rda BBT'nin mevcut uzant\u0131lar\u0131yla a\u00e7\u0131klanaca\u011f\u0131 varsay\u0131lmaktad\u0131r (Khajah, Lindsey ve Mozer, 2016). Bilgiyi izlemeye yakla\u015fmada derin \u00f6\u011frenmenin kullan\u0131m\u0131 verilerde otomatik olarak hala yararl\u0131 ili\u015fkiler bulmaktad\u0131r ancak potansiyel olarak BBT i\u00e7in \u00f6nceden \u00f6nerilen uzant\u0131lara ili\u015fkin ek g\u00f6sterimler bulamam\u0131\u015ft\u0131r. Bu b\u00f6l\u00fcmdeki \u00e7al\u0131\u015fma, \u00f6\u011frencileri temsil etmek i\u00e7in derin a\u011flar\u0131n kullan\u0131lmas\u0131yla ilgilidir ancak yaln\u0131zca de\u011ferlendirme eylemlerinin kullan\u0131lmas\u0131 yerine, her t\u00fcrl\u00fc \u00f6\u011frenci eyleminin dikkate al\u0131nmas\u0131 bak\u0131m\u0131ndan farkl\u0131l\u0131k g\u00f6sterir.<\/span><\/p>\n

\u00d6zellikle, bu b\u00f6l\u00fcmde, n-gram yakla\u015f\u0131m\u0131n\u0131 ve uzun k\u0131sa vadeli haf\u0131za (UKVH) mimarisi olarak bilinen TSA varyant\u0131n\u0131 kullanmay\u0131 d\u00fc\u015f\u00fcnmekteyiz (Hochreiter ve Schmidhuber, 1997). Bu iki model hem veri dizilimini modeller ve hem de sonras\u0131nda hangi i\u015faretin gelmesi gerekti\u011fine dair bir olas\u0131l\u0131k da\u011f\u0131l\u0131m\u0131n\u0131 sa\u011flar. UKVH mimarilerinin ve benzer varyantlar\u0131n kullan\u0131m\u0131 k\u0131smen dizilimlerde uzun -ve k\u0131sa- aral\u0131klardaki ba\u011f\u0131ml\u0131l\u0131klar\u0131n yakalanmas\u0131na izin veren de\u011fi\u015ftirilebilir haf\u0131zas\u0131 sayesinde yak\u0131n zamanda, konu\u015fma, g\u00f6r\u00fcnt\u00fc ve metin analizi de d\u00e2hil olmak \u00fczere s\u0131ral\u0131 verileri i\u00e7eren \u00e7e\u015fitli alanlarda etkileyici sonu\u00e7lar elde etmi\u015ftir (Graves, Mohamed ve Hinton, 2013; Vinyals, Kaiser vd., 2015; Vinyals, Toshev, Bengio ve Erhan, 2015). \u00d6\u011frenci \u00f6\u011frenme davran\u0131\u015f\u0131, sabit bir eylem durum alan\u0131ndaki bir dizi eylem olarak temsil edilebildi\u011finden, ba\u015far\u0131l\u0131 \u00f6\u011frenmeyi karakterize eden karma\u015f\u0131k \u00f6r\u00fcnt\u00fcleri yakalamak i\u00e7in UKVH'ler potansiyel olarak kullan\u0131labilir. \u00d6nceki \u00e7al\u0131\u015fmalarda, \u00f6\u011frenci t\u0131klama verilerinin modellenmesi n-gram modelleri gibi y\u00f6ntemlerle \u00fcmit verici olmu\u015ftur(Wen ve Rose, 2014).<\/span><\/p>\n\n

VER\u0130 K\u00dcMES\u0130<\/h2>\n

Bu b\u00f6l\u00fcmde kullan\u0131lan veri k\u00fcmesi, 2013 Bahar d\u00f6nemi \u0130statistik BerkeleyX KA\u00c7D'den geldi. KA\u00c7D be\u015f hafta boyunca video konferanslar\u0131, ev \u00f6devleri, tart\u0131\u015fma forumlar\u0131 ve iki s\u0131navla devam etti. Orijinal veri k\u00fcmesi, her \u00f6\u011frencinin bir \u015fekilde KA\u00c7D ile etkile\u015fime giren bir kullan\u0131c\u0131 kayd\u0131 oldu\u011fu, 31.000 \u00f6\u011frenciden gelen 17 milyon olay\u0131 i\u00e7ermektedir. Bu etkile\u015fimler, derste belirli bir URL'ye gitme, bir forum mesajlar\u0131nda oylamaya kat\u0131lma, bir s\u0131nav sorusunu cevaplama ve bir konferans videosu oynatma gibi olaylar\u0131 i\u00e7erir. Veriler, her bir kullan\u0131c\u0131n\u0131n t\u00fcm olaylar\u0131n\u0131n s\u0131ralamal\u0131 olarak toplanabilmi\u015f olmas\u0131 i\u00e7in i\u015flenir: 3687 olay t\u00fcr\u00fc m\u00fcmk\u00fcnd\u00fcr. Veri k\u00fcmesindeki her sat\u0131r, ger\u00e7ekle\u015ftirilen eylemi veya \u00f6\u011frenci taraf\u0131ndan eri\u015filen URL'yi temsil eden belirli bir dizine d\u00f6n\u00fc\u015ft\u00fcr\u00fcl\u00fcr.<\/span><\/p>\n

Bu nedenle, her bir kullan\u0131c\u0131n\u0131n i\u015flem grubu, 3687 \u00f6zg\u00fcn t\u00fcr olan bir dizilim indisle temsil edilir. Kay\u0131tl\u0131 etkinlik ge\u00e7mi\u015fimiz, forumun ba\u015fl\u0131klar\u0131, s\u0131navlar, video sayfalar\u0131 ve wiki sayfalar\u0131 gibi dersin farkl\u0131 sayfalar\u0131na giden \u00f6\u011frencileri i\u00e7eriyordu. Bu sayfalarda, video oynatmak ve duraklatmak ya da bir sorunu kontrol etmek gibi sayfa i\u00e7inde yap\u0131lan eylemleri de kaydettik. Ayr\u0131ca s\u0131ral\u0131 olaylar ad\u0131 verilen JavaScript gezinmelerini de kaydederiz. \u00d6n i\u015fleme s\u00fcrecimizin bu a\u00e7\u0131klamas\u0131nda, bu olay dizilimlerini, art arda gelen olay taraf\u0131ndan y\u00f6nlendirilen URL ile a\u00e7\u0131k\u00e7a ili\u015fkilendirilmeden, kendi ba\u015flar\u0131na kaydederiz. Tablo 19.1, veri k\u00fcmesinde mevcut olan farkl\u0131 olay t\u00fcrlerini ve olaya ba\u011fl\u0131 belirli bir URL'yi ili\u015fkilendirmeyi se\u00e7ip se\u00e7medi\u011fimizi listelemektedir. \u00d6n i\u015flemimizde, bu olaylar\u0131n baz\u0131lar\u0131 URL'ye \u00f6zg\u00fc olarak kaydedilir, bu modelin \u00f6\u011frencinin bu olaylar i\u00e7in eri\u015fti\u011fi tam URL\u2019ye maruz kalaca\u011f\u0131 anlam\u0131na gelir. Baz\u0131 olaylar URL'ye \u00f6zg\u00fc olmayan olarak kaydedilir; bu, modelin yaln\u0131zca eylemin ger\u00e7ekle\u015fti\u011fini bildi\u011fi ancak o eylemin derste hangi URL'ye ba\u011fl\u0131 oldu\u011funu bilmedi\u011fi anlam\u0131na gelir. Orijinal veri k\u00fcmesinde 40 kattan daha az ger\u00e7ekle\u015fen olaylar\u0131n ayr\u0131 tutuldu\u011funu da dikkate al\u0131n\u0131z. Bu nedenle, forum etkinliklerinin bir\u00e7o\u011fu URL'ye \u00f6zg\u00fc olduklar\u0131 ancak \u00e7ok s\u0131k ger\u00e7ekle\u015fmedikleri i\u00e7in ayr\u0131 tutulmu\u015flard\u0131r. Git dizilimi, sonraki dizilimi ve \u00f6nceki dizilimi, \u00f6\u011frenciler taray\u0131c\u0131 sayfas\u0131nda g\u00f6r\u00fcnen gezinme d\u00fc\u011fmelerini se\u00e7tiklerinde tetiklenen olaylara at\u0131fta bulunur. Sonraki dizilimi ve \u00f6nceki dizilimi, s\u0131ras\u0131yla dersteki \u00f6nceki veya sonraki i\u00e7erik sayfas\u0131na gider. Git dizilimi, bir b\u00f6l\u00fcm i\u00e7inde bir alt b\u00f6l\u00fcm i\u00e7inden ba\u015fka bir alt b\u00f6l\u00fcme atlamay\u0131 temsil eder.<\/span><\/p>\n

<\/a><\/a><\/a><\/a><\/a> Tablo 19.1. G\u00fcnl\u00fc\u011fe Kaydedilen Olay Tipleri ve \u00d6zellikleri<\/i><\/span><\/span><\/p>\n\n<\/colgroup>\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
Ders Sayfas\u0131 Olaylar\u0131<\/b><\/span><\/span><\/td>\n<\/tr>\n
Sayfa G\u00f6r\u00fcn\u00fcm\u00fc (URL'ye \u00f6zg\u00fc) <\/span><\/span><\/td>\n<\/tr>\n
Git dizilimi (URL'ye \u00f6zg\u00fc de\u011fil) <\/span><\/span><\/td>\n<\/tr>\n
Sonraki dizilimi (URL ye \u00f6zg\u00fc de\u011fil) <\/span><\/span><\/td>\n<\/tr>\n
Dizilim \u00d6nceki (URL' ye \u00f6zg\u00fc de\u011fil)<\/span><\/span><\/td>\n<\/tr>\n
Wiki Olaylar\u0131 <\/b><\/span><\/span><\/td>\n<\/tr>\n
Sayfa G\u00f6r\u00fcn\u00fcm\u00fc (URL'ye \u00f6zg\u00fc) <\/span><\/span><\/td>\n<\/tr>\n
Video Olaylar\u0131<\/b><\/span><\/span><\/td>\n<\/tr>\n
Video Duraklat (URL'ye \u00f6zg\u00fc de\u011fil) <\/span><\/span><\/td>\n<\/tr>\n
Video Oynat\u0131m\u0131 (URL'ye \u00f6zg\u00fc de\u011fil) <\/span><\/span><\/td>\n<\/tr>\n
Problem Olaylar\u0131 <\/b><\/span><\/span><\/td>\n<\/tr>\n
Sayfa G\u00f6r\u00fcn\u00fcm\u00fc (URL'ye \u00f6zg\u00fc de\u011fil) <\/span><\/span><\/td>\n<\/tr>\n
Sorun Kontrolu (URL'ye \u00f6zg\u00fc de\u011fil) <\/span><\/span><\/td>\n<\/tr>\n
Problem Cevap G\u00f6ster(URL'ye \u00f6zg\u00fc de\u011fil)<\/span><\/span><\/td>\n<\/tr>\n
Forum Olaylar\u0131<\/b><\/span><\/span><\/td>\n<\/tr>\n
Forum G\u00f6r\u00fcn\u00fcm\u00fc (URL'ye \u00d6zg\u00fc)<\/span><\/span><\/td>\n<\/tr>\n
Forum Kapat (filtrelenerek ay\u0131r\u0131ld\u0131)<\/span><\/span><\/td>\n<\/tr>\n
Forum Olu\u015ftur (filtrelenerek ayr\u0131ld\u0131 )<\/span><\/span><\/td>\n<\/tr>\n
Forum Silme (filtrelenerek ayr\u0131ld\u0131<\/span><\/span><\/td>\n<\/tr>\n
Forum Teyit (filtrelenerek ayr\u0131ld\u0131)<\/span><\/span><\/td>\n<\/tr>\n
Forum \u0130zlemesi (URL'ye \u00d6zg\u00fc)<\/span><\/span><\/td>\n<\/tr>\n
Forum Yan\u0131t\u0131 (URL'ye \u00d6zg\u00fc)<\/span><\/span><\/td>\n<\/tr>\n
Forum Aramas\u0131 (URL\u2019ye \u00f6zg\u00fc olmayan)<\/span><\/span><\/td>\n<\/tr>\n
Forum Takip Etme (filtrelenerek ayr\u0131ld\u0131)<\/span><\/span><\/td>\n<\/tr>\n
Forum Oy Kullanmama (filtrelenerek ayr\u0131ld\u0131)<\/span><\/span><\/td>\n<\/tr>\n
Forum G\u00fcncellemesi (filtrelenerek ayr\u0131ld\u0131)<\/span><\/span><\/td>\n<\/tr>\n
Forum Be\u011feni (URL'ye \u00d6zg\u00fc)<\/span><\/span><\/td>\n<\/tr>\n
Forum \u0130zlenen Konular\u0131 G\u00f6r\u00fcnt\u00fcle (URL'ye \u00d6zel)<\/span><\/span><\/td>\n<\/tr>\n
Forum Sat\u0131r \u0130\u00e7i G\u00f6r\u00fcnt\u00fcle (URL'ye \u00d6zel)<\/span><\/span><\/td>\n<\/tr>\n
Forum Kullan\u0131c\u0131 Profili G\u00f6r\u00fcnt\u00fcle (URL'ye \u00d6zel)<\/span><\/span><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

\u00d6rne\u011fin, bir \u00f6\u011frenci 2. b\u00f6l\u00fcm 1. alt b\u00f6l\u00fcm URL\u2019sine eri\u015firse, bir ders videosu oynat\u0131rsa, bir sonraki ok d\u00fc\u011fmesine (bir sonraki alt b\u00f6l\u00fcme eri\u015fmek i\u00e7in bir JavaScript gezinmesi ger\u00e7ekle\u015ftirir) t\u0131klar, bir s\u0131nav sorusunu cevaplar ve ard\u0131ndan gezinti \u00e7ubu\u011funda (ba\u015fka bir JavaScript gezintisi ger\u00e7ekle\u015ftiren) 5. alt b\u00f6l\u00fcme t\u0131klar, bu \u00f6\u011frencinin dizilimi be\u015f farkl\u0131 indisle temsil edilir. Birincisi b\u00f6l\u00fcm 2 alt b\u00f6l\u00fcm 1'in URL'sine, ikincisi bir oynatma videosu belirtecine, \u00fc\u00e7\u00fcnc\u00fcs\u00fc bir sonraki gezinti olay\u0131na, d\u00f6rd\u00fcnc\u00fcs\u00fc \u00f6\u011frencinin kursa d\u00e2hil oldu\u011fu belirli probleme gitme olay\u0131 be\u015fincisi gezinti olay\u0131 gite tekab\u00fcl eder. Modele s\u0131rayla bu be\u015f indisin bir listesi verilecek ve sonra neyin gelmesi gerekti\u011fini tahmin etmek i\u00e7in e\u011fitilecektir. Bu nedenle indisler, \u00f6\u011frencinin ger\u00e7ekle\u015ftirdi\u011fi eylem dizilimini temsil eder. Uzunluk i\u00e7in be\u015f gerekli de\u011fildir; \u00fcretici modellere iste\u011fe ba\u011fl\u0131 uzunluktaki dizilimler verilebilir.<\/span><\/p>\n

31.000 \u00f6\u011frenciden, 8094'\u00fc dersin \u00f6\u011fretenleri taraf\u0131ndan \u201conayl\u0131\u201d kabul edilebilecek miktarda \u00f6devi tamamlad\u0131 ve s\u0131navlarda yeterince y\u00fcksek puan ald\u0131lar. Di\u011fer KA\u00c7D ba\u011flamlar\u0131nda, sertifikan\u0131n bazen \u00f6\u011frencinin \u00f6zel bir sertifika i\u00e7in para \u00f6demesi anlam\u0131na geldi\u011fini ancak bu KA\u00c7D i\u00e7in ge\u00e7erli olmad\u0131\u011f\u0131n\u0131 unutmay\u0131n\u0131z. Sertifikal\u0131 \u00f6\u011frenciler, orijinal 17 milyon etkinli\u011fin 11.2 milyonunu olu\u015ftururken, sertifikal\u0131 \u00f6\u011frenci ba\u015f\u0131na ortalama 1390 etkinlik ger\u00e7ekle\u015fti. Sertifikal\u0131 ve sertifikal\u0131 olmayanlar aras\u0131ndaki ayr\u0131m, bu model i\u00e7in \u00f6nemlidir, \u00e7\u00fcnk\u00fc sertifikal\u0131 olarak kabul edilen \u00f6\u011frencilerin eylemlerinin bu KA\u00c7D i\u00e7in makul bir \u015fekilde ba\u015far\u0131l\u0131 bir navigasyon bilgisi \u00f6r\u00fcnt\u00fcs\u00fc olabilece\u011fi hipotezine g\u00f6re \u00fcretici modelleri yaln\u0131zca \"sertifikal\u0131\" kabul edilen \u00f6\u011frencilerin ger\u00e7ekle\u015ftirdi\u011fi eylem dizilimleri \u00fczerine e\u011fitmeyi se\u00e7tik.<\/span><\/p>\n

Veri k\u00fcmesindeki her sat\u0131r, kullan\u0131c\u0131n\u0131n eri\u015fiminin tam URL'si, kullan\u0131c\u0131 i\u00e7in benzersiz bir tan\u0131mlay\u0131c\u0131, i\u015flemin tam olarak ger\u00e7ekle\u015fti\u011fi zaman ve daha fazlas\u0131 gibi i\u015flemle ilgili bilgileri i\u00e7erir. Bu b\u00f6l\u00fcm i\u00e7in, zaman veya muhtemel di\u011fer ba\u011flamsal bilgileri dikkate almay\u0131z, bunun yerine sadece \u00f6\u011frencinin eri\u015fti\u011fi kayna\u011fa veya \u00f6\u011frencinin yapt\u0131\u011f\u0131 eyleme odaklan\u0131r\u0131z. Veri k\u00fcmesinin tamam\u0131nda 40 kattan daha az ger\u00e7ekle\u015fen olaylar, nadiren eri\u015filen tart\u0131\u015fma g\u00f6nderileri veya kullan\u0131c\u0131 profili ziyaretleri olduklar\u0131ndan ve KA\u00c7D'de gezinen di\u011fer \u00f6\u011frencilere uygulanmalar\u0131 muhtemel olmad\u0131\u011f\u0131ndan kald\u0131r\u0131ld\u0131.<\/span><\/p>\n\n

Y\u00d6NTEM<\/h2>\n

Bu \u00e7al\u0131\u015fmada, iki \u00fcretici modelin, tekrarlayan sinir a\u011f\u0131 mimarisi ve n-gram kullan\u0131m\u0131n\u0131 ara\u015ft\u0131rd\u0131k. Bu b\u00f6l\u00fcmde, tekrarlayan sinir a\u011f\u0131n\u0131n mimarisini ve UKVH uzant\u0131s\u0131n\u0131 ayr\u0131nt\u0131lar\u0131yla a\u00e7\u0131kl\u0131yoruz, hipotez olarak sundu\u011fumuz model bir sonraki eylem tahmininde en iyi performans\u0131 g\u00f6sterecektir. N-gram gibi di\u011fer \u201cy\u00fczeysel\u201d modeller daha sonra a\u00e7\u0131klanmaktad\u0131r.<\/span><\/p>\n\n

Tekrarlayan Sinir A\u011flar\u0131<\/h3>\n

Tekrarlayan sinir a\u011flar\u0131 (TSA'lar), keyfi uzunluktaki s\u0131ral\u0131 verileri i\u015flemek i\u00e7in tasarlanm\u0131\u015f bir sinir a\u011f\u0131 modelleri ailesidir. Tekrarlayan sinir a\u011flar\u0131, belirli bir diziliminin i\u015flenmesi s\u00fcresince devam eden s\u00fcrekli ve gizli bir durumu etraf\u0131nda tutarak \u00e7al\u0131\u015f\u0131r. Bu gizli durum, \u015fimdiye kadarki dizilim ile ilgili bilgileri yakalar, b\u00f6ylece dizimin sonraki b\u00f6l\u00fcmlerindeki \u00f6ng\u00f6r\u00fc, bu s\u00fcrekli gizli durumdan etkilenebilir. Ad\u0131ndan da anla\u015f\u0131laca\u011f\u0131 gibi, TSA'lar ileri beslemeli sinir a\u011flar\u0131 taraf\u0131ndan kullan\u0131lan hesaplama yakla\u015f\u0131m\u0131n\u0131 kullan\u0131r ve ayn\u0131 zamanda zaman ad\u0131mlar\u0131 aras\u0131nda devam eden s\u00fcrekli bir gizli durumu da dayat\u0131r. Gizli durumu bir giri\u015f dizilimindeki elemanlar aras\u0131nda tutmak, tekrarlayan sinir a\u011flar\u0131na s\u0131ral\u0131 modelleme g\u00fcc\u00fc veren \u015feydir. Bu \u00e7al\u0131\u015fmada, TSA'ya her girdi KA\u00c7D veri k\u00fcmesinden gelen gran\u00fcl bir \u00f6\u011frenci olay\u0131 olacakt\u0131r. TSA, \u015fimdiye kadar g\u00f6r\u00fclen olaylara dayanarak \u00f6\u011frencilerin bir sonraki olay\u0131n\u0131 tahmin etmek i\u00e7in e\u011fitilmi\u015ftir. \u015eekil 19.1, girdilerin \u00f6\u011frencilerin eylemleri olaca\u011f\u0131 ve \u00e7\u0131kt\u0131lar\u0131n dizilim i\u00e7inden bir sonraki \u00f6\u011frenci hareketi olaca\u011f\u0131 basit bir TSA diyagram\u0131n\u0131 g\u00f6sterir. A\u015fa\u011f\u0131daki denklemler, TSA modelinin parametrelerinin her birinde kullan\u0131lan matematiksel i\u015flemleri g\u00f6sterir: ht, s\u00fcrekli gizli durumu temsil eder. Bu gizli durum, xt + 1'deki tahminin gizli durum ht'den etkilenece\u011fi \u015feklinde bulundurulur. TSA modeli, bir giri\u015f a\u011f\u0131rl\u0131\u011f\u0131 matrisi Wx, tekrarlayan a\u011f\u0131rl\u0131k matrisi Wh, ba\u015flang\u0131\u00e7 durumu h0 ve \u00e7\u0131k\u0131\u015f matrisi Wy ile parametrelendirilir: bh ile by s\u0131ras\u0131yla gizli ve \u00e7\u0131k\u0131\u015f birimleri i\u00e7in sapmalard\u0131r.<\/span><\/p>\n

ht<\/sub> = tanh(W x<\/sub>xt<\/sub> + W h<\/sub>ht\u22121<\/sub> + bh<\/sub>) (1) <\/span><\/p>\n

yt<\/sub> =\u03c3(W y<\/sub>ht<\/sub> +by<\/sub>) (2)<\/span><\/p>\n

\"\"<\/p>\n<\/a> \u015eekil 19.1. Basit tekrarlayan sinir a\u011f\u0131<\/i><\/span><\/span>\n

UKVH Modelleri<\/h3>\n

TSA'n\u0131n pop\u00fcler bir \u00e7e\u015fidi, gizli durumda ne zaman anlaml\u0131 bilgiler tutaca\u011f\u0131n\u0131 ve gizli durumu ne zaman temizleyece\u011fi veya ne zaman \"unutaca\u011f\u0131n\u0131\" \u00f6\u011frenen \u201ckap\u0131lar\u201d ekleyerek TSA\u2019lar\u0131n uzun d\u00f6nem ba\u011f\u0131ml\u0131l\u0131klar\u0131 \u00f6\u011frenmelerine yard\u0131mc\u0131 oldu\u011fu d\u00fc\u015f\u00fcn\u00fclen, anlaml\u0131 uzun vadeli etkile\u015fimlerin s\u00fcreklili\u011fine izin veren uzun k\u0131sa s\u00fcreli bellek (UKVH; Hochreiter ve Schmidhuber, 1997) mimarisidir. UKVH'ler, gizli durumun ne zaman temizlenece\u011fini ve ne zaman yararl\u0131 bilgilerle g\u00fc\u00e7lenece\u011fini belirlemek i\u00e7in a\u00e7\u0131k\u00e7a \u00f6\u011frenilen ek ge\u00e7it parametreleri ekler. Bunun yerine, her gizli durum, h1<\/sub> ek ge\u00e7it parametreleri i\u00e7eren bir UKVH h\u00fccre birimi ile de\u011fi\u015ftirilir. Bu kap\u0131lar nedeniyle, UKVH'lerin basit TSA'lardan daha etkili bir \u015fekilde e\u011fitildi\u011fi bulunmu\u015ftur (Bengio, Simard ve Frasconi, 1994; Gers, Schmidhuber ve Cummins, 2000). Bir UKVH i\u00e7in g\u00fcncelleme denklemleri a\u015fa\u011f\u0131daki gibidir:<\/span><\/p>\nf<\/i><\/span>t<\/span><\/sub> = <\/span>\u03c3<\/span>(<\/span>W <\/i><\/span>fx<\/span><\/sub>x<\/i><\/span>t <\/span><\/sub>+ <\/span>W <\/i><\/span>fh<\/span><\/sub>h<\/i><\/span>t <\/span><\/sub>\u2212 1<\/i><\/span><\/sub> <\/i>+ <\/span>b<\/i><\/span>f<\/span><\/sub>) (3) <\/span><\/span><\/span>\n\ni<\/i><\/span>t<\/span><\/sub> = \u03c3(<\/span>W<\/i><\/span>ix<\/span><\/sub>x<\/i><\/span>t<\/span><\/sub> <\/i>+ <\/span>W<\/i><\/span>ih<\/span><\/sub>h<\/i><\/span>t<\/span><\/sub>\u22121 <\/i><\/span><\/sub>+ <\/span>b<\/i><\/span>i<\/span><\/sub>) (4) <\/span><\/span><\/span>\n\nC<\/i><\/span>~<\/i><\/span><\/sup>t<\/i><\/span><\/sub> <\/i>= <\/span>tanh<\/i><\/span>(<\/span>W<\/i><\/span>Cx<\/i><\/span><\/sub>x<\/i><\/span>t<\/i><\/span><\/sub> <\/i>+ <\/span>W <\/i><\/span>Ch<\/i><\/span><\/sub>h<\/i><\/span>t\u22121<\/i><\/span><\/sub> +<\/span>b<\/i><\/span>C<\/i><\/span><\/sub>) <\/span>(5)<\/span> <\/span><\/span>\n\nCt <\/i><\/span>= <\/span>ft <\/i><\/span>\u00d7 <\/span>Ct<\/i><\/span>\u22121 + <\/span>it <\/i><\/span>\u00d7 <\/span>C\u02dc<\/i><\/span>t<\/i><\/span><\/sub> <\/i>(6) <\/span><\/span><\/span>\n\not <\/i><\/span>= <\/span>\u03c3<\/span> (<\/span>W <\/i><\/span>ox<\/i><\/span><\/sub>x<\/i><\/span>t<\/i><\/span><\/sub> <\/i>+ <\/span>W <\/i><\/span>oh<\/i><\/span><\/sub>h<\/i><\/span>t<\/i><\/span><\/sub>\u22121 <\/span><\/sub>+ <\/span>b<\/i><\/span>o<\/i><\/span><\/sub>) <\/span>(7) <\/span><\/span><\/span>\n\nht <\/i><\/span>= <\/span>o<\/i><\/span>t<\/i><\/span><\/sub> <\/i>\u00d7 <\/span>tanh<\/i><\/span>(<\/span>Ct<\/i><\/span>) (8)<\/span><\/span><\/span>\n

\"\"<\/p>\n<\/a> \u015eekil 19.2. UKVH i\u00e7in g\u00fcncelleme denklemlerine kar\u015f\u0131l\u0131k gelen say\u0131larla bir h\u00fccrenin anatomisi.<\/i><\/span><\/span>\n

\u015eekil 19.2, \u015fekildeki say\u0131lar\u0131n UKVH: ft, it ve ot i\u00e7in, daha \u00f6nce bahsedilen g\u00fcncelleme denklemlerine kar\u015f\u0131l\u0131k geldi\u011fi bir h\u00fccrenin anatomisini g\u00f6sterir; bu, UKVH taraf\u0131ndan \u00f6nceki h\u00fccreden \"unutma\" verilerini ve yeni h\u00fccre durumuna neyin \"girilece\u011fini\" ve h\u00fccre durumundan neyin \"\u00e7\u0131kaca\u011f\u0131n\u0131\" belirlemek i\u00e7in kullan\u0131lan ge\u00e7it mekanizmalar\u0131n\u0131 temsil eder. Ct, bilgilerin UKVH'yi beslemesi s\u0131ras\u0131nda bilgilerin \u00e7\u0131kar\u0131ld\u0131\u011f\u0131 ve eklendi\u011fi gizli h\u00fccre durumunu temsil eder. C\u02dct, bir sonraki h\u00fccre durumunu g\u00fcncellemek i\u00e7in ge\u00e7itli hale getirilmi\u015f olan yeni aday h\u00fccre durumunu temsil eder.<\/span><\/p>\n\n

UKVH Uygulamas\u0131<\/h3>\n

Bu b\u00f6l\u00fcmde kullan\u0131lan \u00fcretici UKVH modelleri, Theano'nun \u00fczerine in\u015fa edilmi\u015f bir Python k\u00fct\u00fcphanesi olan Keras (Chollet, 2015) kullan\u0131larak ger\u00e7ekle\u015ftirilmi\u015ftir (Bergstra vd., 2010; Bastien vd., 2012). Model, bir indeks numaras\u0131 ile temsil edilen her \u00f6\u011frenci eylemini ger\u00e7ekle\u015ftirir. Bu indisler, yapay de\u011fi\u015fkenlik olarak da bilinen 1-s\u0131cak vekt\u00f6r kodlamas\u0131ndaki indekse kar\u015f\u0131l\u0131k gelir. Model her dizini bir g\u00f6mme vekt\u00f6r\u00fcne d\u00f6n\u00fc\u015ft\u00fcr\u00fcr ve sonra g\u00f6m\u00fcl\u00fc vekt\u00f6r\u00fc birer birer t\u00fcketir. G\u00f6mme katman\u0131n\u0131n kullan\u0131m\u0131 s\u00f6zc\u00fckleri \u00e7ok boyutlu semantik bir uzaya e\u015flemenin bir yolu olarak do\u011fal dil i\u015fleme s\u00fcre\u00e7lerinde ve dil modellemede yayg\u0131nd\u0131r (Goldberg ve Levy, 2014). Burada bir g\u00f6mme katman\u0131, KA\u00c7D eylem alan\u0131ndaki eylemler i\u00e7in benzer bir e\u015flemenin olabilece\u011fi hipotezi ile birlikte kullan\u0131lmaktad\u0131r. Model, daha \u00f6nce \u00f6\u011frenci taraf\u0131ndan ger\u00e7ekle\u015ftirilen eylemler g\u00f6z \u00f6n\u00fcne al\u0131nd\u0131\u011f\u0131nda, bir sonraki \u00f6\u011frenci eylemini tahmin etmek i\u00e7in e\u011fitilmi\u015ftir. Zaman i\u00e7inde geri yay\u0131l\u0131m (Werbos, 1988), UKVH parametrelerini e\u011fitmek i\u00e7in kullan\u0131l\u0131r, bir sonraki eylemin indeksi olan bir ger\u00e7ek referans de\u011fer olarak softmax katman\u0131 kullan\u0131l\u0131r. Kay\u0131p hesaplan\u0131rken kategorik \u00e7apraz entropi, RMSprop ise en iyile\u015ftirici olarak kullan\u0131l\u0131r. UKVH katmanlar\u0131 aras\u0131na a\u015f\u0131r\u0131 uyumu engelleme y\u00f6ntemi olarak b\u0131rakma katmanlar\u0131 eklenmi\u015ftir (Pham, Bluche, Kermorvant ve Louradour, 2014). Her bir e\u011fitim verisi grubu i\u00e7in rastgele s\u0131f\u0131rlama y\u00fczdesi a\u011f kenar\u0131 a\u011f\u0131rl\u0131klar\u0131n\u0131n belirli bir y\u00fczdelik grubu s\u0131f\u0131rlar. Gelecekteki \u00e7al\u0131\u015fmalarda, \u00f6zellikle UKVH'ler ve TSA'lar i\u00e7in haz\u0131rlanm\u0131\u015f di\u011fer d\u00fczenlile\u015ftirme tekniklerini de\u011ferlendirmek faydal\u0131 olabilir (Zaremba, Sutskever ve Vinyals, 2014). Yaln\u0131zca veri dizisindeki gezinme eylemlerini \u00e7\u0131karmakla ba\u015flayan \u00f6n i\u015fleme ve UKVH model kodumuzun bir versiyonunu kamuya a\u00e7\u0131k hale getirdik4<\/a><\/sup>.<\/span><\/p>\n\n

UKVH Hiperparametre Aramas\u0131<\/h3>\n

\u0130lk ara\u015ft\u0131rmam\u0131z\u0131n bir par\u00e7as\u0131 olarak 24 UKVH modelinin her birini farkl\u0131 hiperparametre k\u00fcmeleri ile 10 evrede5<\/a><\/sup> e\u011fittik. Bir evre, veriler aras\u0131nda tam bir ge\u00e7i\u015f yapan parametre yerle\u015ftirme algoritmas\u0131d\u0131r. UKVH modellerimiz i\u00e7in aranan hiperparametre alan\u0131 Tablo 19.2'de g\u00f6sterilmektedir. Bu hiperparametreler, etki b\u00fcy\u00fckl\u00fc\u011f\u00fcne g\u00f6re farkl\u0131 hiperparametrelere \u00f6ncelik veren \u00f6nceki \u00e7al\u0131\u015fmalara dayanarak \u015febeke aramalar\u0131 i\u00e7in se\u00e7ilmi\u015ftir (Greff, Srivastava, Koutnik, Steunebrink ve Schmidhuber, 2015). Zamanlama ad\u0131na, 3 katl\u0131 UKVH modellerini.0001 \u00f6\u011frenme oranlar\u0131 ile e\u011fitmemeyi tercih ettik. Ayr\u0131ca, ek hiperparametreyi ve e\u011fitim y\u00f6ntemlerini ara\u015ft\u0131rmak i\u00e7in ba\u015flang\u0131\u00e7 noktas\u0131 olarak hizmet etmek \u00fczere ilk incelemenin sonu\u00e7lar\u0131n\u0131 kulland\u0131\u011f\u0131m\u0131z geni\u015fletilmi\u015f bir ara\u015ft\u0131rma yapt\u0131k.<\/span><\/p>\n

E\u011fitim TSA'lar\u0131 nispeten zaman al\u0131c\u0131 oldu\u011fundan, geni\u015fletilmi\u015f ara\u015ft\u0131rma, \u00fcmit vaat eden hiperparametre kombinasyonlar\u0131n\u0131n bir alt k\u00fcmesinden olu\u015fuyordu (Sonu\u00e7lar b\u00f6l\u00fcm\u00fcne bak\u0131n).<\/span><\/p>\n\n

    \n \t
  1. \n
      \n \t
    1. \n
        \n \t
      1. \n
          \n \t
        1. \n

          <\/a><\/a><\/a><\/a><\/a> Tablo 19.2. UKVH Hiperparametre G\u00f6zene\u011fi<\/i><\/span><\/span><\/p>\n<\/li>\n<\/ol>\n<\/li>\n<\/ol>\n<\/li>\n<\/ol>\n<\/li>\n<\/ol>\n<\/colgroup>\n\n\n\n\n
          \n

          Gizli Katmanlar<\/b><\/span><\/p>\n<\/td>\n

          		\n

          1 <\/b><\/span><\/p>\n<\/td>\n

          		\n

          2 <\/b><\/span><\/p>\n<\/td>\n

          		\n

          3<\/b><\/span><\/p>\n<\/td>\n<\/tr>\n

          \n

          Gizli Katmandaki D\u00fc\u011f\u00fcmler<\/b><\/span><\/p>\n<\/td>\n

          		\n

          64 <\/span><\/p>\n<\/td>\n

          		\n

          128 <\/span><\/p>\n<\/td>\n

          		\n

          256<\/span><\/p>\n<\/td>\n<\/tr>\n

          \n

          \u00d6\u011frenme oran\u0131 (_)<\/b><\/span><\/p>\n<\/td>\n

          		\n

          .001<\/span><\/p>\n<\/td>\n

          		\n

          .0001 <\/span><\/p>\n<\/td>\n

          		\n

          .0001*<\/span><\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

          \u00c7apraz Do\u011frulama<\/h3>\n

          Her modelin yorday\u0131c\u0131 g\u00fcc\u00fcn\u00fc de\u011ferlendirmek i\u00e7in 5 katl\u0131 \u00e7apraz do\u011frulama kullan\u0131lm\u0131\u015ft\u0131r. Her model verinin %80'i i\u00e7in e\u011fitildi ve sonra kalan %20'si \u00fczerinde do\u011fruland\u0131; Bu be\u015f kez yap\u0131ld\u0131, b\u00f6ylece her \u00f6\u011frenci eylemi bir kez bir do\u011frulama setine girdi. UKVH'ler i\u00e7in, model e\u011fitim s\u00fcrecinde do\u011frulama kesinli\u011fi hakk\u0131nda bilgi sa\u011flamak amac\u0131yla t\u0131rmanma seti olarak hizmet vermek i\u00e7in haz\u0131rl\u0131k verilerinin %10'unu ger\u00e7ekle\u015ftirdi. D\u00fczenlenen setteki her sat\u0131r bir \u00f6\u011frencinin ald\u0131\u011f\u0131 t\u00fcm eylem dizilimlerinden olu\u015fur. Model taraf\u0131ndan \u00fcretilen sonraki do\u011fru eylem tahminlerinin oran\u0131, her \u00f6\u011frenci eylem dizilimi i\u00e7in hesaplan\u0131r. S\u00f6z konusu k\u0131vr\u0131ma6<\/a><\/sup> y\u00f6nelik modelin performans\u0131n\u0131 \u00fcretmek i\u00e7in b\u00fct\u00fcn bir k\u0131vr\u0131m\u0131n oranlar\u0131n\u0131n ortalamas\u0131 al\u0131n\u0131r ve daha sonra belirli bir UKVH model hiperparametre seti i\u00e7in \u00c7D do\u011frulu\u011funu \u00fcretmek \u00fczere be\u015f k\u0131vr\u0131m\u0131n tamam\u0131ndaki performans\u0131n ortalamas\u0131 al\u0131n\u0131r.<\/span><\/p>\n\n

          Y\u00fczeysel Modeller<\/h3>\n

          N-gram modeller, basit ama g\u00fc\u00e7l\u00fc olas\u0131l\u0131kl\u0131 modellerdir ve dizilimlerin yap\u0131s\u0131n\u0131 gram olarak adland\u0131r\u0131lan n-boyutlu alt dizilerin istatistikleri yoluyla yakalamay\u0131 hedefler ve n-s\u0131ral\u0131 Markov zincirlerine e\u015fde\u011ferdir. Spesifik olarak, model, xi'nin e\u011fitim setinde \u00f6nceki n-1 durumlar\u0131n\u0131 takip etme olas\u0131l\u0131\u011f\u0131 olan tahmini ko\u015fullu olas\u0131l\u0131k P(xi<\/sub>|i<\/sub>xi<\/sub>_(n_1)<\/sub>, xi\u20131<\/sub>), kullanarak her bir dizi durumunu tahmin eder. N-gram modeller hem h\u0131zl\u0131 ve basit hesaplan\u0131r ve do\u011frudan yorumlara sahiptir. Eylem alan\u0131ndaki olas\u0131 her eylem i\u00e7in bir parametre atayan nispeten y\u00fcksek parametre modelleri olduklar\u0131ndan, n gamlar\u0131n olduk\u00e7a rekabet\u00e7i bir standart olmas\u0131n\u0131 bekliyoruz.<\/span><\/p>\n

          N-gram modellerinde, n'nin 2 ila 10 aras\u0131nda de\u011fi\u015fti\u011fini de\u011ferlendirdik; bunlar\u0131n en b\u00fcy\u00fc\u011f\u00fc, e\u011fitim s\u0131ras\u0131nda UKVH ba\u011flam penceresinin boyutuna kar\u015f\u0131l\u0131k gelmektedir. E\u011fitim setinin hi\u00e7bir g\u00f6zlem i\u00e7ermeyen tahminlerini ele almak i\u00e7in, en az bir g\u00f6zlem i\u00e7eren en b\u00fcy\u00fck n-gram\u0131n tahminine tekrar tekrar d\u00f6nmeye dayanan bir y\u00f6ntem olan gerilemeyi7<\/a><\/sup> kulland\u0131k. Do\u011frulama stratejimiz, UKVH modelleriyle ayn\u0131yd\u0131, burada ayn\u0131 be\u015f kat\u0131n ortalama \u00e7apraz do\u011frulama puan\u0131 her model i\u00e7in hesapland\u0131.<\/span><\/p>\n\n

          Ders Yap\u0131s\u0131 Modelleri<\/h3>\n

          Dizlim verilerinin varsay\u0131msal yap\u0131sal \u00f6zelliklerinden yararlanmaya y\u00f6nelik \u00e7e\u015fitli alternatif modelleri de d\u00e2hil ettik. Dizilimleri incelerken fark etti\u011fimiz ilk \u015fey, belirli eylemlerin arka arkaya birka\u00e7 kez tekrarlanmas\u0131yd\u0131. Bu nedenle, bu varsay\u0131m\u0131n tek ba\u015f\u0131na veri k\u00fcmesindeki bir sonraki eylemi ne kadar iyi tahmin edebilece\u011fini bilmek \u00f6nemlidir. Daha sonra, ders i\u00e7eri\u011fi en s\u0131k sabit bir dizilimde d\u00fczenlendi\u011finden, ders izlencesinin bir sonraki sayfay\u0131 veya eylemi tahmin etme yetene\u011fini de\u011ferlendirdik. Bunu ders i\u00e7eri\u011findeki sayfalar\u0131, eylem setimizdeki \u00f6\u011frenci sayfa ge\u00e7i\u015fleri ile e\u015fle\u015ftirerek, ders izlencesindeki toplam 300 maddeden 174'\u00fcn\u00fcn e\u015fle\u015fmesi ile sonu\u00e7land\u0131rarak ba\u015fard\u0131k. Eylem alan\u0131m\u0131zda her zaman bulunmayan i\u00e7erik kimli\u011fi dizgilerini e\u015fle\u015ftirmeye g\u00fcvendi\u011fimizden, k\u00fc\u00e7\u00fck bir \u00fcst \u00fcste binen eylemler alt k\u00fcmesi e\u015fle\u015ftirilmedi. Son olarak, mevcut durumun ders program\u0131 i\u00e7inde olmamas\u0131 durumunda, mevcut durumun bir sonraki durum olarak \u00f6ng\u00f6r\u00fclmesi y\u00f6n\u00fcnden her iki modeli de birle\u015ftirdik.<\/span><\/p>\n\n

          SONU\u00c7LAR<\/h2>\n

          Bu b\u00f6l\u00fcmde, farkl\u0131 \u00f6\u011frenme oranlar\u0131, katman ba\u015f\u0131na gizli d\u00fc\u011f\u00fcm say\u0131s\u0131 ve UKVH katman say\u0131s\u0131 ile e\u011fitilmi\u015f, daha \u00f6nce bahsedilen UKVH modellerinin sonu\u00e7lar\u0131n\u0131 tart\u0131\u015f\u0131yoruz. Model ba\u015far\u0131s\u0131, 5 kat \u00e7apraz do\u011frulama ile belirlenir ve modelin bir sonraki eylemi ne kadar iyi tahmin etti\u011fi ile ilintilidir. N-gram modelleri ve di\u011fer rota yap\u0131s\u0131 modelleri, 5 kat \u00e7apraz do\u011frulama ile do\u011frulan\u0131r.<\/span><\/p>\n\n

          UKVH Modelleri<\/h3>\n

          Tablo 19.3, 10 yeti\u015ftirme s\u00fcresinden sonra hesaplanan t\u00fcm 24 UKVH modelinin \u00c7D do\u011frulu\u011funu g\u00f6stermektedir. .01 \u00f6\u011frenme h\u0131z\u0131na sahip modeller i\u00e7in, tepe t\u0131rmanma setlerinde do\u011fruluk genellikle yineleme 10'da zirveye ula\u015ft\u0131. D\u00fc\u015f\u00fck \u00f6\u011frenme oranlar\u0131na sahip modeller i\u00e7in, en y\u00fcksek \u00c7D do\u011fruluklar\u0131n\u0131n daha fazla e\u011fitimle geli\u015fmesini beklemek makul olacakt\u0131r. Bu modellerin e\u011fitim s\u00fcrecinde ne kadar iyi performans g\u00f6sterdi\u011fine dair bir anl\u0131k g\u00f6r\u00fcnt\u00fc sa\u011flamak yerine 10 tekrardan sonra sonu\u00e7lar\u0131 basit\u00e7e rapor etmeyi se\u00e7tik. Ayr\u0131ca, uzun vadede model performans\u0131n\u0131n.01 \u00f6\u011frenme oran\u0131 model performanslar\u0131 \u00fczerinde ciddi oranda bir geli\u015fme g\u00f6sterme ihtimalinin olmad\u0131\u011f\u0131n\u0131 ve s\u0131n\u0131rl\u0131 GPU hesaplama kaynaklar\u0131nda \u00e7al\u0131\u015facak en umut verici ke\u015fifleri en \u00fcst d\u00fczeye \u00e7\u0131karmam\u0131z gerekti\u011fini varsay\u0131yoruz. Her \u00f6\u011frenme oran\u0131 i\u00e7in en iyi \u00c7D do\u011frulu\u011fu vurgulamak i\u00e7in koyu renkli yap\u0131lm\u0131\u015ft\u0131r.<\/span><\/p>\n

          UKVH'leri kullanman\u0131n bir dezavantaj\u0131, bir GPU'ya ihtiya\u00e7 duymalar\u0131 ve e\u011fitilmelerinin nispeten yava\u015f olmas\u0131d\u0131r. Bu nedenle, kullan\u0131lacak en iyi hiperparametreleri ara\u015ft\u0131r\u0131rken, yaln\u0131zca ilk ke\u015fiflerin bir alt k\u00fcmesini temel alan ek modeller e\u011fitmeyi se\u00e7tik. Ayr\u0131ca, ge\u00e7mi\u015f ba\u011flam\u0131 10 \u00f6geden 100 \u00f6geye geni\u015fleterek modele maruz kalan ba\u011flam\u0131 artt\u0131rd\u0131k. Tablo 4, bu geni\u015fletilmi\u015f sonu\u00e7lar\u0131 g\u00f6stermektedir. Her UKVH katman\u0131 256 d\u00fc\u011f\u00fcme sahiptir ve \u00f6nceki hiperparametre arama sonu\u00e7lar\u0131ndaki 10 evre yerine, 20 veya 60 evre i\u00e7in e\u011fitilmi\u015ftir. Geni\u015fletilmi\u015f sonu\u00e7lar, \u00f6nceki sonu\u00e7lara g\u00f6re b\u00fcy\u00fck bir iyile\u015fme g\u00f6stermekte olup, yeni do\u011fruluk, .7093'e k\u0131yasla .7223'te zirveye ula\u015fm\u0131\u015ft\u0131r.<\/span><\/p>\n

          \u015eekil 19.3, ilk ke\u015fiften itibaren 1 ve 2 katmanl\u0131 modeller i\u00e7in evreli e\u011fitim s\u0131ras\u0131ndaki %10 tepe t\u0131rmanma hold out k\u00fcmesinin do\u011frulama kesinli\u011fini g\u00f6stermektedir. Her veri noktas\u0131, belirli bir katman ve d\u00fc\u011f\u00fcm say\u0131s\u0131 kombinasyonu i\u00e7in her \u00fc\u00e7 \u00f6\u011frenme h\u0131z\u0131ndaki ortalama t\u0131rmanma do\u011frulu\u011funu temsil eder. Ampirik olarak, daha fazla say\u0131da d\u00fc\u011f\u00fcme sahip olmak, ilk 10 evrede daha y\u00fcksek bir do\u011frulukla ili\u015fkilendirilirken, 2 katmanl\u0131 modeller, kar\u015f\u0131l\u0131k gelen 1 katmanl\u0131 modele yakla\u015fmadan veya ondan \u00f6nce, birka\u00e7 evre i\u00e7in d\u00fc\u015f\u00fck do\u011frulama kesinlikleriyle ba\u015flar. Bu rakam ilk 10 evre i\u00e7in bir anl\u0131k g\u00f6r\u00fcnt\u00fc sa\u011flar; a\u00e7\u0131k\u00e7as\u0131 baz\u0131 parametre kombinasyonlar\u0131 i\u00e7in, daha fazla evre, ek geni\u015fletilmi\u015f UKVH aramas\u0131yla g\u00f6sterildi\u011fi gibi daha y\u00fcksek bir tepe t\u0131rmanma do\u011frulu\u011funa neden olacakt\u0131r. Tahmini olarak, 3 katmanl\u0131 modeller de 2 katmanl\u0131 modellerin sergiledi\u011fi, do\u011fruluklar\u0131n daha alt katman emsallerine k\u0131yasla ba\u015flang\u0131\u00e7ta daha d\u00fc\u015f\u00fck ba\u015flayabildi\u011fi bir e\u011filimi izleyebilir.<\/span><\/p>\n

          <\/a><\/a><\/a><\/a><\/a> Tablo 19.3. UKVH Performans\u0131 (10 Evre)<\/i><\/span><\/span><\/p>\n\n<\/colgroup>\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
          \n

          \u00d6\u011frenme Oran\u0131<\/span><\/p>\n<\/td>\n

          		\n

          D\u00fc\u011f\u00fcmler<\/span><\/p>\n<\/td>\n

          		\n

          Katmanlar<\/span><\/p>\n<\/td>\n

          		\n

          Do\u011fruluk<\/span><\/p>\n<\/td>\n<\/tr>\n<\/thead>\n

          0.01<\/span><\/span><\/td>\n64<\/span><\/span><\/td>\n1<\/span><\/span><\/td>\n0.7014<\/span><\/span><\/td>\n<\/tr>\n
          0.01<\/span><\/span><\/td>\n64<\/span><\/span><\/td>\n2<\/span><\/span><\/td>\n0.7009<\/span><\/span><\/td>\n<\/tr>\n
          0.01<\/span><\/span><\/td>\n64<\/span><\/span><\/td>\n3<\/span><\/span><\/td>\n0.6997<\/span><\/span><\/td>\n<\/tr>\n
          0.01<\/span><\/span><\/td>\n128<\/span><\/span><\/td>\n1<\/span><\/span><\/td>\n0.7046<\/span><\/span><\/td>\n<\/tr>\n
          0.01<\/span><\/span><\/td>\n128<\/span><\/span><\/td>\n2<\/span><\/span><\/td>\n0.7064<\/span><\/span><\/td>\n<\/tr>\n
          0.01<\/span><\/span><\/td>\n128<\/span><\/span><\/td>\n3<\/span><\/span><\/td>\n0.7056<\/span><\/span><\/td>\n<\/tr>\n
          0.01<\/span><\/span><\/td>\n256<\/span><\/span><\/td>\n1<\/span><\/span><\/td>\n0.7073<\/span><\/span><\/td>\n<\/tr>\n
          0.01<\/span><\/span><\/td>\n256<\/span><\/span><\/td>\n2<\/span><\/span><\/td>\n0.7093<\/span><\/span><\/td>\n<\/tr>\n
          0.01<\/span><\/span><\/td>\n256<\/span><\/span><\/td>\n3<\/span><\/span><\/td>\n0.7092<\/span><\/span><\/td>\n<\/tr>\n
          0.001<\/span><\/span><\/td>\n64<\/span><\/span><\/td>\n1<\/span><\/span><\/td>\n0.6941<\/span><\/span><\/td>\n<\/tr>\n
          0.001<\/span><\/span><\/td>\n64<\/span><\/span><\/td>\n2<\/span><\/span><\/td>\n0.6968<\/span><\/span><\/td>\n<\/tr>\n
          0.001<\/span><\/span><\/td>\n64<\/span><\/span><\/td>\n3<\/span><\/span><\/td>\n0.6971<\/span><\/span><\/td>\n<\/tr>\n
          0.001<\/span><\/span><\/td>\n128<\/span><\/span><\/td>\n1<\/span><\/span><\/td>\n0.6994<\/span><\/span><\/td>\n<\/tr>\n
          0.001<\/span><\/span><\/td>\n128<\/span><\/span><\/td>\n2<\/span><\/span><\/td>\n0.7022<\/span><\/span><\/td>\n<\/tr>\n
          0.001<\/span><\/span><\/td>\n128<\/span><\/span><\/td>\n3<\/span><\/span><\/td>\n0.7026<\/span><\/span><\/td>\n<\/tr>\n
          0.001<\/span><\/span><\/td>\n256<\/span><\/span><\/td>\n1<\/span><\/span><\/td>\n0.7004<\/span><\/span><\/td>\n<\/tr>\n
          0.001<\/span><\/span><\/td>\n256<\/span><\/span><\/td>\n2<\/span><\/span><\/td>\n0.7050<\/span><\/span><\/td>\n<\/tr>\n
          0.001<\/span><\/span><\/td>\n256<\/span><\/span><\/td>\n3<\/span><\/span><\/td>\n0.7050<\/span><\/span><\/td>\n<\/tr>\n
          0.0001<\/span><\/span><\/td>\n64<\/span><\/span><\/td>\n1<\/span><\/span><\/td>\n0.6401<\/span><\/span><\/td>\n<\/tr>\n
          0.0001<\/span><\/span><\/td>\n64<\/span><\/span><\/td>\n2<\/span><\/span><\/td>\n0.4719<\/span><\/span><\/td>\n<\/tr>\n
          0.0001<\/span><\/span><\/td>\n128<\/span><\/span><\/td>\n1<\/span><\/span><\/td>\n0.6539<\/span><\/span><\/td>\n<\/tr>\n
          0.0001<\/span><\/span><\/td>\n128<\/span><\/span><\/td>\n2<\/span><\/span><\/td>\n0.6648<\/span><\/span><\/td>\n<\/tr>\n
          0.0001<\/span><\/span><\/td>\n256<\/span><\/span><\/td>\n1<\/span><\/span><\/td>\n0.6677<\/span><\/span><\/td>\n<\/tr>\n
          0.0001<\/span><\/span><\/td>\n256<\/span><\/span><\/td>\n2<\/span><\/span><\/td>\n0.6894<\/span><\/span><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n \n

          <\/a><\/a><\/a><\/a><\/a> Tablo 19.4. Geni\u015fletilmi\u015f UKVH Performans\u0131 (256 D\u00fc\u011f\u00fcm, 100 Pencere Boyutu) <\/i><\/span><\/span><\/p>\n\n<\/colgroup>\n\n\n\n\n\n\n\n\n\n\n
          \n

          \u00d6\u011frenme Oran\u0131<\/span><\/p>\n<\/td>\n

          		\n

          Evre<\/span><\/p>\n<\/td>\n

          		\n

          Katmanlar<\/span><\/p>\n<\/td>\n

          		\n

          Do\u011fruluk<\/span><\/p>\n<\/td>\n<\/tr>\n

          0.01<\/span><\/span><\/td>\n20<\/span><\/span><\/td>\n2<\/span><\/span><\/td>\n0.7190<\/span><\/span><\/td>\n<\/tr>\n
          0.01<\/span><\/span><\/td>\n60<\/span><\/span><\/td>\n2<\/span><\/span><\/td>\n0.7220<\/span><\/span><\/td>\n<\/tr>\n
          0.01<\/span><\/span><\/td>\n20<\/span><\/span><\/td>\n3<\/span><\/span><\/td>\n0.7174<\/span><\/span><\/td>\n<\/tr>\n
          0.01<\/span><\/span><\/td>\n60<\/span><\/span><\/td>\n3<\/span><\/span><\/td>\n0.7223<\/span><\/span><\/td>\n<\/tr>\n
          0.001<\/span><\/span><\/td>\n20<\/span><\/span><\/td>\n2<\/span><\/span><\/td>\n0.7044<\/span><\/span><\/td>\n<\/tr>\n
          0.001<\/span><\/span><\/td>\n60<\/span><\/span><\/td>\n2<\/span><\/span><\/td>\n0.7145<\/span><\/span><\/td>\n<\/tr>\n
          0.001<\/span><\/span><\/td>\n20<\/span><\/span><\/td>\n3<\/span><\/span><\/td>\n0.7039<\/span><\/span><\/td>\n<\/tr>\n
          0.001<\/span><\/span><\/td>\n60<\/span><\/span><\/td>\n3<\/span><\/span><\/td>\n0.7147<\/span><\/span><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

          Ders Yap\u0131s\u0131 Modelleri<\/h3>\n

          Farkl\u0131 ders yap\u0131 modelleri i\u00e7in model performans\u0131 Tablo 19.5'te g\u00f6sterilmi\u015ftir. Sonu\u00e7lar, UKVH veya n-gram sonu\u00e7lar\u0131 aral\u0131\u011f\u0131nda dura\u011fanl\u0131k (sonuncusuyla ayn\u0131) veya ders i\u00e7eri\u011fi yap\u0131s\u0131 gibi basit sezgisel taramalardan ya da her iki bulu\u015fsal y\u00f6ntemi de i\u00e7eren (\"ders program\u0131 + tekrar\") bir\u00e7ok eylemin tahmin edilebilece\u011fini d\u00fc\u015f\u00fcnd\u00fcrmektedir.<\/span><\/p>\n

          \"\"<\/p>\n<\/a> \u015eekil 19.3. Her bir e\u011fitim k\u00fcmesinin %10'unu olu\u015fturan tepe t\u0131rmanma verilerinde evreye g\u00f6re ortalama do\u011fruluk.<\/i><\/span><\/span>\n

          <\/a><\/a><\/a><\/a><\/a> Tablo 19.5. Yap\u0131sal modeller<\/i><\/span><\/span><\/p>\n\n<\/colgroup>\n\n\n\n\n\n
          \n

          Yap\u0131sal model<\/b><\/span><\/p>\n<\/td>\n

          		\n

          Do\u011fruluk<\/b><\/span><\/p>\n<\/td>\n<\/tr>\n

          tekrarlama<\/span><\/span><\/td>\n0.2908<\/span><\/span><\/td>\n<\/tr>\n
          \u00f6\u011fretim izlencesi<\/span><\/span><\/td>\n0.2339<\/span><\/span><\/td>\n<\/tr>\n
          \u00f6\u011fretim izlencesi + tekrar<\/span><\/span><\/td>\n0.4533<\/span><\/span><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n
            \n \t
          1. \n
              \n \t
            1. \n
                \n \t
              1. \n
                  \n \t
                1. \n

                  N-gram Modelleri<\/h4>\n<\/li>\n<\/ol>\n<\/li>\n<\/ol>\n<\/li>\n<\/ol>\n<\/li>\n<\/ol>\n

                  Model performans\u0131 Tablo 19.6'da g\u00f6sterilmi\u015ftir. En iyi performans g\u00f6steren modeller, \u00f6nceki 7 veya 8 eylemi kullanarak (s\u0131ras\u0131yla 8 ve 9 gram) tahminlerde bulundu. Performans\u0131 art\u0131rmayan daha geni\u015f kay\u0131tlar n aral\u0131\u011f\u0131m\u0131z\u0131n yeterince b\u00fcy\u00fck oldu\u011funu g\u00f6stermi\u015ftir. Genel olarak performans, en iyi n-gram modelin en iyi UKVH modellerinden daha k\u00f6t\u00fc \u00e7al\u0131\u015fmas\u0131na ra\u011fmen n-gram modellerinin UKVH modelleriyle rekabet etti\u011fini g\u00f6stermektedir. Tablo 19.7, en karma\u015f\u0131k model (10 gram) i\u00e7in kullan\u0131lan n-gram modellerin oran\u0131n\u0131 g\u00f6stermektedir. Tahminlerin %62'sinden fazlas\u0131, 10 graml\u0131k g\u00f6zlemler kullan\u0131larak yap\u0131lm\u0131\u015ft\u0131r. Ayr\u0131ca, vakalar\u0131n %1'inden az\u0131 tahminleri yapmak i\u00e7in unigramlara veya bigramlara geri d\u00f6nd\u00fc ve bu da daha b\u00fcy\u00fck gram \u00f6r\u00fcnt\u00fcleri i\u00e7in \u00f6nemli bir g\u00f6zlem eksikli\u011fi olmad\u0131\u011f\u0131n\u0131 \u00f6ne s\u00fcrd\u00fc.<\/span><\/p>\n

                  <\/a><\/a><\/a><\/a><\/a> Tablo 19.6. N-gram Performans<\/i><\/span><\/span><\/p>\n\n<\/colgroup>\n\n\n\n\n\n\n\n\n\n\n\n
                  \n

                  N-gram<\/span><\/p>\n<\/td>\n

                  		\n

                  Do\u011fruluk<\/span><\/p>\n<\/td>\n<\/tr>\n

                  2-gram<\/span><\/span><\/td>\n0.6304<\/span><\/span><\/td>\n<\/tr>\n
                  3 gram<\/span><\/span><\/td>\n0.6658<\/span><\/span><\/td>\n<\/tr>\n
                  4 gram<\/span><\/span><\/td>\n0.6893<\/span><\/span><\/td>\n<\/tr>\n
                  5 gram<\/span><\/span><\/td>\n0.6969<\/span><\/span><\/td>\n<\/tr>\n
                  6 gram<\/span><\/span><\/td>\n0.7012<\/span><\/span><\/td>\n<\/tr>\n
                  7 gram<\/span><\/span><\/td>\n0.7030<\/span><\/span><\/td>\n<\/tr>\n
                  8 gram<\/span><\/span><\/td>\n0.7035<\/span><\/span><\/td>\n<\/tr>\n
                  9 gram<\/span><\/span><\/td>\n0.7035<\/span><\/span><\/td>\n<\/tr>\n
                  10 gram<\/span><\/span><\/td>\n0.7033<\/span><\/span><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n \n

                  <\/a><\/a><\/a><\/a><\/a> Tablo 19.7. N cinsinden 10 graml\u0131k \u00f6ng\u00f6r\u00fcn\u00fcn oran\u0131<\/i><\/span><\/span><\/p>\n\n<\/colgroup>\n\n\n\n\n\n\n\n\n\n\n\n\n\n
                  \n

                  n<\/span><\/p>\n<\/td>\n

                  		\n

                  \u00d6ng\u00f6r\u00fclen %<\/b><\/span><\/p>\n<\/td>\n<\/tr>\n<\/thead>\n

                  1 <\/span><\/span><\/td>\n0.0003<\/span><\/span><\/td>\n<\/tr>\n
                  2<\/span><\/span><\/td>\n0.0084<\/span><\/span><\/td>\n<\/tr>\n
                  3<\/span><\/span><\/td>\n0.0210<\/span><\/span><\/td>\n<\/tr>\n
                  4<\/span><\/span><\/td>\n0.0423<\/span><\/span><\/td>\n<\/tr>\n
                  5<\/span><\/span><\/td>\n0.0524<\/span><\/span><\/td>\n<\/tr>\n
                  6<\/span><\/span><\/td>\n0.0605<\/span><\/span><\/td>\n<\/tr>\n
                  7<\/span><\/span><\/td>\n0.0624<\/span><\/span><\/td>\n<\/tr>\n
                  8<\/span><\/span><\/td>\n0.0615<\/span><\/span><\/td>\n<\/tr>\n
                  9<\/span><\/span><\/td>\n0.0594<\/span><\/span><\/td>\n<\/tr>\n
                  10<\/span><\/span><\/td>\n0.6229<\/span><\/span><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

                  Yine de yakla\u015f\u0131k %6 daha az veri noktas\u0131 art arda gelen daha b\u00fcy\u00fck n-gramlar taraf\u0131ndan tahmin ediliyor gibi g\u00f6r\u00fcnmektedir.<\/span><\/p>\n\n

                    \n \t
                  1. \n
                      \n \t
                    1. \n
                        \n \t
                      1. \n
                          \n \t
                        1. \n

                          Sertifikas\u0131z \u00d6\u011frencilerin Do\u011frulanmas\u0131<\/h4>\n<\/li>\n<\/ol>\n<\/li>\n<\/ol>\n<\/li>\n<\/ol>\n<\/li>\n<\/ol>\n

                          Sonunda sertifika almayan \u00f6\u011frencilerden gelen veri ak\u0131\u015flar\u0131 \u00fczerindeki eylemleri tahmin etmek i\u00e7in 10 evreli e\u011fitimden sonra (.01 \u00f6\u011frenme h\u0131z\u0131, 256 d\u00fc\u011f\u00fcm, iki katman) en iyi performans g\u00f6steren \u201corijinal\u201d UKVH modelini kulland\u0131k. Sertifikas\u0131z \u00f6\u011frencilerin \u00e7o\u011fu yaln\u0131zca birka\u00e7 oturum a\u00e7ma eylemi ger\u00e7ekle\u015ftirdi, bu nedenle analizi en az 30 oturum a\u00e7ma eylemi olan \u00f6\u011frencilerle s\u0131n\u0131rlad\u0131k. Bu kriterleri kar\u015f\u0131layan 10.761 \u00f6\u011frenci ve toplam 2.151.666 eylem vard\u0131. UKVH modeli, sertifikal\u0131 \u00f6\u011frenciler i\u00e7in .7093 \u00e7apraz do\u011frulanm\u0131\u015f do\u011frulukla kar\u015f\u0131la\u015ft\u0131r\u0131ld\u0131\u011f\u0131nda, do\u011frulanmam\u0131\u015f \u00f6\u011frenci alan\u0131ndan gelen eylemleri .6709 do\u011frulukla tam bir \u015fekilde tahmin edebildi. Bu fark, sertifikal\u0131 \u00f6\u011frencilerden gelen eylemlerin, belgelendirilmemi\u015f \u00f6\u011frencilerden gelen eylemlerden farkl\u0131 olma e\u011filiminde oldu\u011funu g\u00f6sterir, belki de \u00f6\u011frencilere rehberlik etmek i\u00e7in otomatik bir \u00f6neri \u00e7er\u00e7evesi sa\u011flamada potansiyel bir uygulama g\u00f6stermektedir.<\/span><\/p>\n

                          <\/a> Tablo 19.8. \u00c7apraz Do\u011frulanm\u0131\u015f Modellerin Kar\u015f\u0131la\u015ft\u0131r\u0131lmas\u0131<\/i><\/span><\/span><\/p>\n\n<\/colgroup>\n\n\n\n\n
                          \n

                          N-gram<\/b><\/span><\/p>\n<\/td>\n

                          		\n

                          Do\u011fru<\/b><\/span><\/p>\n<\/td>\n

                          		\n

                          N-gram Yanl\u0131\u015f<\/b><\/span><\/p>\n<\/td>\n<\/tr>\n

                          UKVH Do\u011fru<\/span><\/span><\/td>\n7.565.862<\/span><\/span><\/td>\n577.683<\/span><\/span><\/td>\n<\/tr>\n
                          UKVH Yanl\u0131\u015f<\/span><\/span><\/td>\n367.960<\/span><\/span><\/td>\n2.735.702<\/span><\/span><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

                          KATKILAR<\/h2>\n

                          Bu \u00e7al\u0131\u015fmada, gran\u00fcler \u00f6\u011frenci eylem verilerinin modellenmesi sorununa bir KA\u00c7D i\u00e7indeki her t\u00fcr etkile\u015fimi modelleyerek yakla\u015ft\u0131k. Bu \u00f6ncelikli olarak de\u011ferlendirme sonu\u00e7lar\u0131n\u0131 kullanarak gizli \u00f6\u011frenci bilgisini modellemeye odaklanan \u00f6nceki \u00e7al\u0131\u015fmalardan farkl\u0131d\u0131r. Bir \u00f6\u011frencinin bir sonraki eylemini tahmin ederken, en iyi performans g\u00f6steren UKVH modeli .7223 \u00e7apraz do\u011frulama kesinli\u011fi \u00fcretti ki bu da en iyi n-gram model do\u011frulu\u011fu olan .7035'in \u00fczerinde bir geli\u015fme olarak toplam 11 milyon olas\u0131 tahminden 210.000 daha do\u011fru tahmindir. Tablo 19.8, iki modelin \u00e7apraz onaylama s\u0131ras\u0131nda do\u011fru veya yanl\u0131\u015f bir \u00f6ng\u00f6r\u00fcde mutab\u0131k kald\u0131\u011f\u0131 veya kalmad\u0131\u011f\u0131 say\u0131y\u0131 g\u00f6stermektedir. Hem UKVH hem de n-gram modelleri, bir sonraki eylemin \u00f6\u011fretim program\u0131 izlencesi yap\u0131s\u0131 ve tekrarlar arac\u0131l\u0131\u011f\u0131yla \u00f6ng\u00f6r\u00fclmesi yap\u0131sal modeli \u00fczerinde \u00f6nemli bir geli\u015fme sa\u011flar; bu, \u00f6\u011frenci kat\u0131l\u0131m \u00f6r\u00fcnt\u00fclerinin ders materyali i\u00e7erisinde tamamen do\u011frusal bir gezinmeden a\u00e7\u0131k\u00e7a sapt\u0131\u011f\u0131n\u0131 g\u00f6sterir.<\/span><\/p>\n

                          Bildi\u011fimiz kadar\u0131yla, bu b\u00f6l\u00fcm bir KA\u00c7D'de davran\u0131\u015fsal verinin bu ayr\u0131nt\u0131 d\u00fczeyinde tahmin edildi\u011fini ilk kez g\u00f6stermektedir. Ayr\u0131ca, KA\u00c7D verilerine ilk defa tekrarlayan sinir a\u011flar\u0131n\u0131n uyguland\u0131\u011f\u0131n\u0131 g\u00f6sterir. Bu tekni\u011fin \u00f6\u011frenci davran\u0131\u015fsal durumlar\u0131n\u0131 ham zaman serisi verileri ile temsil etmek i\u00e7in, \u00f6zellik m\u00fchendisli\u011fi olmadan, y\u00fcksek hacimli zaman serisi verileri ile herhangi bir \u00f6\u011frenme analiti\u011fi ba\u011flam\u0131nda geni\u015f bir uygulanabilirli\u011fe sahip oldu\u011funa inan\u0131yoruz. \u00c7er\u00e7evelememiz, davran\u0131\u015fsal veri modellerinin \u00f6\u011frenci i\u00e7in gelecekteki davran\u0131\u015flar\u0131 \u00f6nermek i\u00e7in nas\u0131l kullan\u0131labilece\u011fini ortaya koyarken, davran\u0131\u015fsal durumlar\u0131n\u0131n temsili, performanstan duyu\u015fsal duruma kadar \u00e7e\u015fitli yap\u0131larda \u00e7e\u015fitli \u00e7\u0131kar\u0131mlar yapmak i\u00e7in de\u011ferli olabilir.<\/span><\/p>\n\n

                          GELECEKTE YAPILACAKLAR<\/h2>\n

                          Hem UKVH hem de n-gram modelleri de geli\u015ftirilebilir. \u00d6zellikle, n-gram modellerimiz g\u00f6r\u00fcnmeyen gramlar\u0131n daha iyi kullan\u0131lmas\u0131na izin veren geri tepme ve p\u00fcr\u00fczs\u00fczle\u015ftirme tekniklerinin bir kombinasyonundan faydalanabilir. UKVH, daha geni\u015f bir hiperparametre g\u00f6zene\u011fi arama, daha fazla yeti\u015ftirme s\u00fcresi, daha uzun yeti\u015ftirme ba\u011flam\u0131 pencereleri ve daha y\u00fcksek boyutlu eylem yerle\u015ftirmelerinden faydalanabilir. Ek olarak, veri k\u00fcmemizdeki sinyal-g\u00fcr\u00fclt\u00fc oran\u0131, daha az bilgilendirici veya gereksiz \u00f6\u011frenci eylemleri kald\u0131r\u0131larak veya eylemler aras\u0131ndaki s\u00fcreyi temsil etmek i\u00e7in ek belirte\u00e7ler eklenerek artt\u0131r\u0131labilir.<\/span><\/p>\n

                          Derin \u00f6\u011frenme modellerini b\u00fcy\u00fck \u00f6\u011frenci eylemi veri k\u00fcmelerine uygulaman\u0131n birincil nedeni, KA\u00c7D ortamlar\u0131ndaki \u00f6\u011frenci davran\u0131\u015f\u0131n\u0131 modellemektir. Bu \u00f6r\u00fcnt\u00fcler, otomatikle\u015ftirilmi\u015f \u00f6neri sistemlerinin olu\u015fturulmas\u0131na yard\u0131mc\u0131 olmak i\u00e7in kullan\u0131labilir; burada, zorlu bir \u00f6\u011frenciye, ge\u00e7mi\u015f davran\u0131\u015flar\u0131na ve performanslar\u0131na g\u00f6re i\u00e7eri\u011fi g\u00f6r\u00fcnt\u00fclemek i\u00e7in ge\u00e7i\u015f \u00f6nerileri sa\u011flanabilir. B\u00f6yle bir uygulaman\u0131n olas\u0131l\u0131\u011f\u0131n\u0131 de\u011ferlendirmek i\u00e7in, a\u011f\u0131m\u0131zdan t\u00fcretilmi\u015f bir \u00f6neri sistemini, y\u00f6nlendirilmemi\u015f bir kontrol grubuna kar\u015f\u0131 deneysel olarak test etmeyi planl\u0131yoruz. Ek olarak, gelecekteki \u00e7al\u0131\u015fmalar, \u00e7e\u015fitli dersler i\u00e7in benzer modellerin performans\u0131n\u0131 de\u011ferlendirmeli ve tek bir model kullanarak genel ders \u00f6r\u00fcnt\u00fclerinin ne \u00f6l\u00e7\u00fcde \u00f6\u011frenilebilece\u011fini incelemelidir. Bu b\u00f6l\u00fcmde \u00f6nerilen modeller, bilgi i\u015flemsel bir davran\u0131\u015f modelini s\u00fcrd\u00fcrmektedir. KA\u00c7D'lerdeki \u00f6\u011frenci davran\u0131\u015f dizilimlerinde d\u00fczenliliklerin var oldu\u011fu bu modeller arac\u0131l\u0131\u011f\u0131yla g\u00f6sterilmi\u015ftir. Bilgi i\u015flemsel bir modelin bu kal\u0131plar\u0131 saptayabildi\u011fi g\u00f6z \u00f6n\u00fcne al\u0131nd\u0131\u011f\u0131nda, model bize \u00f6\u011frenci davran\u0131\u015flar\u0131 hakk\u0131nda daha geni\u015f kapsaml\u0131 ne s\u00f6yleyebilir ve bu bulgular mevcut davran\u0131\u015f teorileriyle nas\u0131l ba\u011flant\u0131 kurabilir ve bunlar\u0131 nas\u0131l kurabilir? Her zaman diliminde model \u00f6\u011frenci i\u00e7in gizli bir davran\u0131\u015f durumunu takip etti\u011finden, bu durum o anda mevcut oldu\u011fu bilinen \u00f6\u011frencilerin di\u011fer \u00f6zellikleri ile g\u00f6rselle\u015ftirilebilir ve ili\u015fkilendirilebilir. Gelecekteki \u00e7al\u0131\u015fmalar, \u00f6\u011frencinin durumu hakk\u0131ndaki kendi anlay\u0131\u015f\u0131m\u0131z\u0131 bilgilendirmeye yard\u0131mc\u0131 olabilmesi i\u00e7in bu bilgi i\u015flemsel davran\u0131\u015f modelini geli\u015ftirmeye \u00e7al\u0131\u015facakt\u0131r.<\/span><\/p>\n\n

                          TE\u015eEKK\u00dcR B\u00d6L\u00dcM\u00dc<\/h2>\n

                          Bu \u00e7al\u0131\u015fma, Ulusal Bilim Vakf\u0131'ndan bir hibe ile desteklenmi\u015ftir (IIS: BIGDATA 1547055).<\/span><\/p>\n\n

                          KAYNAK\u00c7A<\/h2>\nBastien, F., Lamblin, P., Pascanu, R., Bergstra, J., Goodfellow, I. J., Bergeron, A., . . . Bengio, Y. (2012). Theano: New features and speed improvements. Deep Learning and Unsupervised Feature Learning NIPS 2012 Workshop. Advances in Neural Information Processing Systems 25 <\/i>(NIPS 2012), 3\u20138 December 2012, Lake Tahoe, NV, USA. http:\/\/www.iro.umontreal.ca\/~lisa\/pointeurs\/nips2012_deep_workshop_theano_final.pdf <\/span><\/span>\n\nBengio, Y., Simard, P., & Frasconi, P. (1994). Learning long-term dependencies with gradient descent is difficult. IEEE Transactions on Neural Networks, 5<\/i>(2), 157\u2013166. <\/span><\/span>\n\nBergstra, J., Breuleux, O., Bastien, F., Lamblin, P., Pascanu, R., Desjardins, G., . . . Bengio, Y. (2010, June). Theano: A CPU and GPU math expression compiler. Proceedings of the Python for Scientific Computing Conference <\/i>(SciPy 2010), 28 June\u20133 July 2010, Austin, TX, USA (pp. 3\u201310).<\/span><\/span>\n\nBrown, P. F., Desouza, P. V., Mercer, R. L., Pietra, V. J. D., & Lai, J. C. (1992). Class-based n-gram models of natural language. Computational Linguistics, 18<\/i>(4), 467\u2013479. <\/span><\/span>\n\nChollet, F. (2015). Keras. GitHub. https:\/\/github.com\/fchollet\/keras <\/span><\/span>\n\nCorbett, A. T., & Anderson, J. R. (1994). Knowledge tracing: Modeling the acquisition of procedural knowledge. User Modeling and User-Adapted Interaction, 4<\/i>(4), 253\u2013278. <\/span><\/span>\n\nCrossley, S., Paquette, L., Dascalu, M., McNamara, D. S., & Baker, R. S. (2016). Combining click-stream data with NLP tools to better understand MOOC completion. Proceedings of the 6th International Conference on Learning Analytics and Knowledge <\/i>(LAK\u201916), 25\u201329 April 2016, Edinburgh, UK (pp. 6\u201314). New York: ACM. <\/span><\/span>\n\nGers, F. A., Schmidhuber, J., & Cummins, F. (2000). Learning to forget: Continual prediction with LSTM. Neural Computation, 12<\/i>(10), 2451\u20132471. <\/span><\/span>\n\nGoldberg, Y., & Levy, O. (2014). Word2vec explained: Deriving Mikolov et al.\u2019s negative-sampling word-embedding method. CoRR. arxiv.org\/abs\/1402.3722 <\/span><\/span>\n\nGraves, A., Mohamed, A.-r., & Hinton, G. (2013). Speech recognition with deep recurrent neural networks. Proceedings of the 2013 IEEE International Conference on Acoustics, Speech and Signal Processing <\/i>(ICASSP 2013), 26\u201331 May, Vancouver, BC, Canada (pp. 6645\u20136649). Institute of Electrical and Electronics Engineers. <\/span><\/span>\n\nGreff, K., Srivastava, R. K., Koutn\u00edk, J., Steunebrink, B. R., & Schmidhuber, J. (2015). LSTM: A search space odyssey. arXiv preprint arXiv:1503.04069. <\/span><\/span>\n\nHochreiter, S., & Schmidhuber, J. (1997). Long short-term memory. Neural Computation, 9<\/i>(8), 1735\u20131780. <\/span><\/span>\n\nKhajah, M., Lindsey, R. V., & Mozer, M. C. (2016). How deep is knowledge tracing? arXiv preprint arXiv:1604.02416. <\/span><\/span>\n\nMikolov, T., Karafi\u00e1t, M., Burget, L., Cernocky, J., & Khudanpur, S. (2010). Recurrent neural network based language model. Proceedings of the 11th Annual Conference of the International Speech Communication Association <\/i>(INTERSPEECH 2010), 26\u201330 September 2010, Makuhari, Chiba, Japan (pp. 1045\u20131048). http:\/\/www.fit.vutbr.cz\/research\/groups\/speech\/publi\/2010\/mikolov_interspeech2010_IS100722.pdf <\/span><\/span>\n\nOleksandra, P., & Shane, D. (2016). Untangling MOOC learner networks. Proceedings of the 6th International Conference on Learning Analytics and Knowledge <\/i>(LAK\u201916), 25\u201329 April 2016, Edinburgh, UK (pp. 208\u2013212). New York: ACM. <\/span><\/span>\n\nPardos, Z. A., Bergner, Y., Seaton, D. T., & Pritchard, D. E. (2013). Adapting Bayesian knowledge tracing to a massive open online course in EDX. In S. K. D\u2019Mello et al. (Eds.), Proceedings of the 6th International Conference on Educational Data Mining <\/i>(EDM2013), 6\u20139 July 2013, Memphis, TN, USA (pp. 137\u2013144). International Educational Data Mining Society\/Springer. <\/span><\/span>\n\nPardos, Z. A., & Xu, Y. (2016). Improving efficacy attribution in a self-directed learning environment using prior knowledge individualization. Proceedings of the 6th International Conference on Learning Analytics and Knowledge <\/i>(LAK\u201916), 25\u201329 April 2016, Edinburgh, UK (pp. 435\u2013439). New York: ACM. <\/span><\/span>\n\nPham, V., Bluche, T., Kermorvant, C., & Louradour, J. (2014). Dropout improves recurrent neural networks for handwriting recognition. Proceedings of the 14th International Conference on Frontiers in Handwriting Recognition <\/i>(ICFHR 2014) 1\u20134 September 2014, Crete, Greece (pp. 285\u2013290). <\/span><\/span>\n\nPiech, C., Bassen, J., Huang, J., Ganguli, S., Sahami, M., Guibas, L. J., & Sohl-Dickstein, J. (2015). Deep knowledge tracing. In C. Cortes et al. (Eds.), Advances in Neural Information Processing Systems 28 <\/i>(NIPS 2015), 7\u201312 December 2015, Montreal, QC, Canada (pp. 505\u2013513). <\/span><\/span>\n\nReddy, S., Labutov, I., & Joachims, T. (2016). Latent skill embedding for personalized lesson sequence recommendation. CoRR. arxiv.org\/abs\/1602.07029 <\/span><\/span>\n\nReich, J., Stewart, B., Mavon, K., & Tingley, D. (2016). The civic mission of MOOCs: Measuring engagement across political differences in forums. Proceedings of the 3rd ACM Conference on Learning @ Scale <\/i>(L@S 2016), 25\u201328 April 2016, Edinburgh, Scotland (pp. 1\u201310). New York: ACM.<\/span><\/span>\n\nSharma, A., Biswas, A., Gandhi, A., Patil, S., & Deshmukh, O. (2016). Livelinet: A multimodal deep recurrent neural network to predict liveliness in educational videos. In T. Barnes et al. (Eds.), Proceedings of the 9th International Conference on Educational Data Mining <\/i>(EDM2016), 29 June\u20132 July 2016, Raleigh, NC, USA. International Educational Data Mining Society. http:\/\/www.educationaldatamining.org\/EDM2016\/proceedings\/paper_64.pdf <\/span><\/span>\n\nVinyals, O., Kaiser, L. Koo, T., Petrov, S., Sutskever, I., & Hinton, G. (2015). Grammar as a foreign language. In C. Cortes et al. (Eds.), Advances in Neural Information Processing Systems 28 <\/i>(NIPS 2015), 7\u201312 December 2015, Montreal, QC, Canada (pp. 2755\u20132763). http:\/\/papers.nips.cc\/paper\/5635-grammar-as-a-foreign-language.pdf <\/span><\/span>\n\nVinyals, O., Toshev, A., Bengio, S., & Erhan, D. (2015). Show and tell: A neural image caption generator. Proceedings of the 2015 IEEE Computer Society Conference on Computer Vision and Pattern Recognition <\/i>(CVPR 2015), 8\u201310 June 2015, Boston, MA, USA. IEEE Computer Society. arXiv:1411.4555 <\/span><\/span>\n\nWen, M., & Ros\u00e9, C. P. (2014). Identifying latent study habits by mining learner behavior patterns in massive open online courses. Proceedings of the 23rd ACM International Conference on Information and Knowledge Management <\/i>(CIKM\u201914), 3\u20137 November 2014, Shanghai, China (pp. 1983\u20131986). New York: ACM. <\/span><\/span>\n\nWen, M., Yang, D., & Ros\u00e9, C. P. (2014). Sentiment analysis in MOOC discussion forums: What does it tell us? In J. Stamper et al. (Eds.), Proceedings of the 7th International Conference on Educational Data Mining <\/i>(EDM2014), 4\u20137 July 2014, London, UK. International Educational Data Mining Society. http:\/\/www.cs.cmu.edu\/~mwen\/papers\/edm2014-camera-ready.pdf <\/span><\/span>\n\nWerbos, P. J. (1988). Generalization of backpropagation with application to a recurrent gas market model. Neural Networks, 1<\/i>(4), 339\u2013356. <\/span><\/span>\n\nZaremba, W., Sutskever, I., & Vinyals, O. (2014). Recurrent neural network regularization. arXiv:1409.2329.<\/span><\/span>\n\n
                          \n\n
                          \n

                          1<\/a> orj. feature engineering<\/span><\/span><\/p>\n\n<\/div>\n

                          \n

                          2<\/a>orj. generative<\/span><\/span><\/p>\n\n<\/div>\n

                          \n

                          3<\/a>orj. embedding<\/span><\/span><\/p>\n\n<\/div>\n

                          \n

                          4<\/a> https:\/\/github.com\/CAHLR\/mooc-behaviorcase-study<\/span><\/span><\/p>\n\n<\/div>\n

                          \n

                          5<\/a> orj. epoch<\/span><\/span><\/p>\n\n<\/div>\n

                          \n

                          6<\/a> orj. fold<\/span><\/span><\/p>\n\n<\/div>\n

                          \n

                          7<\/a> orj.backoff<\/span><\/span><\/p>\n\n<\/div>\n","rendered":"

                          Steven Tang, Joshua C. Peterson ve Zachary A. Pardos<\/span><\/span><\/p>\n

                          E\u011fitim Bilimleri Enstit\u00fcs\u00fc, UC Berkeley, ABD<\/span><\/span><\/p>\n

                          DOI: 10.18608\/hla17.019<\/span><\/span><\/p>\n

                          \u00d6Z<\/h2>\n

                          Kitlesel a\u00e7\u0131k \u00e7evrimi\u00e7i dersler (KA\u00c7D), \u00f6\u011frenenlerin \u00f6\u011frenme materyalleriyle etkile\u015fime girerken ve kavrama etkinliklerini tamamlarken g\u00f6sterdikleri eylemlerin kapsaml\u0131 bir kayd\u0131n\u0131 olu\u015fturur. Bu y\u00fcksek hacimli s\u0131ral\u0131 veri ve se\u00e7im ile \u00f6\u011frenci davran\u0131\u015f\u0131n\u0131 modelleme potansiyeli gelmektedir. \u00d6\u011frenme ortamlar\u0131ndan kaydedilenler gibi uzun vadeli, s\u0131ral\u0131 verilere bakmak i\u00e7in \u00e7e\u015fitli y\u00f6ntemler vard\u0131r. Dil modellemesi alan\u0131nda, geleneksel n-gram teknikleri ve modern tekrarlayan sinir a\u011flar\u0131 (TSA) yakla\u015f\u0131mlar\u0131, dilde yap\u0131y\u0131 algoritmik olarak bulmak ve \u00f6nceki kelimeleri girdi olarak verilen c\u00fcmle veya paragrafta bir sonraki s\u00f6zc\u00fc\u011f\u00fc tahmin etmek i\u00e7in uygulanmaktad\u0131r. Bu b\u00f6l\u00fcmde biz bu \u00e7al\u0131\u015fmaya, bir KA\u00c7D’deki kaynak g\u00f6r\u00fcn\u00fcmleri ve etkile\u015fimlerinin \u00f6\u011frenci dizilimlerini girdiler olarak ele alarak ve \u00f6\u011frencilerin bir sonraki etkile\u015fimini \u00e7\u0131kt\u0131lar olarak tahmin ederek bir benzetim yap\u0131yoruz. <\/span><\/span><\/p>\n

                          Bizim yakla\u015f\u0131m\u0131m\u0131z, KA\u00c7D’de kaynaklar\u0131n temsilini belirgin bir \u00f6zellik m\u00fchendisli\u011fi1<\/a><\/sup> gerektirmeden \u00f6\u011frenir. Bu model potansiyel olarak, bir \u00f6\u011frencinin ba\u015far\u0131 elde etmek i\u00e7in bir sonraki ad\u0131mda yapmas\u0131 gereken eylemlere dair \u00f6neriler \u00fcretmek i\u00e7in kullan\u0131labilir. Ek olarak, b\u00f6yle bir model otomatik olarak performans ve duyu\u015f hakk\u0131nda \u00e7\u0131kar\u0131m sa\u011flayan bir \u00f6\u011frenci davran\u0131\u015fsal durumu olu\u015fturur. \u00c7al\u0131\u015fmam\u0131zda kullan\u0131lan KA\u00c7D\u2019nin 3.500\u2019den fazla e\u015fsiz kayna\u011f\u0131 oldu\u011fu g\u00f6z \u00f6n\u00fcne al\u0131nd\u0131\u011f\u0131nda, bir \u00f6\u011frencinin etkile\u015fime girece\u011fi bir sonraki kesin kayna\u011f\u0131 tahmin etmek zor bir s\u0131n\u0131fland\u0131rma problemi gibi g\u00f6r\u00fcnebilir. Ders program\u0131n\u0131n (dersin yap\u0131s\u0131n\u0131n) bu \u00f6ng\u00f6r\u00fcy\u00fc yapmada ortalama %23 do\u011fruluk sa\u011flad\u0131\u011f\u0131n\u0131, ard\u0131ndan %70.4 ile n-gram y\u00f6ntemini ve %72.2 ile TSA bazl\u0131 y\u00f6ntemleri takip etti\u011fini ke\u015ffettik. Bu ara\u015ft\u0131rma, \u00f6zellik m\u00fchendisli\u011fi gerektirmeyen teknikler kullanarak ince taneli zaman serisi \u00f6\u011frenci verilerinin davran\u0131\u015f modellemesi i\u00e7in zemin haz\u0131rlamaktad\u0131r.<\/span><\/span><\/p>\n

                          Anahtar Kelimeler<\/span>: Davran\u0131\u015f modellemesi, dizilim tahmini, KA\u00c7D’ler, TSA’lar<\/span><\/span><\/p>\n

                          G\u00fcn\u00fcm\u00fcz\u00fcn dijital d\u00fcnyas\u0131, b\u00fcy\u00fck kullan\u0131c\u0131 eylemlerinin g\u00fcnl\u00fck kay\u0131tlar\u0131na dayanan ki\u015fiselle\u015ftirme ile i\u015faretlenmi\u015ftir. E\u011fitim alan\u0131nda, kullan\u0131c\u0131n\u0131n (genellikle gizli) \u00f6zelliklerine g\u00f6re \u00f6\u011frenme \u00f6nerilerini ve sonu\u00e7lar\u0131n\u0131 bireysel kullan\u0131c\u0131lara uyarlayabilen ki\u015fiselle\u015ftirilmi\u015f ve otomatik \u00f6\u011freticilere y\u00f6nelik ara\u015ft\u0131rmalar devam etmektedir. Son y\u0131llarda, kitlesel a\u00e7\u0131k \u00e7evrimi\u00e7i dersler (KA\u00c7D’ler) gibi y\u00fcksek\u00f6\u011frenim \u00e7evrimi\u00e7i \u00f6\u011frenme ortamlar\u0131 \u00f6\u011frenci taraf\u0131ndan olu\u015fturulan y\u00fcksek miktarda \u00f6\u011frenme eylemlerini bir araya getirmi\u015ftir. Bu b\u00f6l\u00fcmde, \u00f6\u011frenmeyi istendi\u011fi kadar eri\u015filebilir, sa\u011flam ve verimli k\u0131lmak i\u00e7in \u00f6\u011frenme yollar\u0131n\u0131 ki\u015fiselle\u015ftirme becerisine y\u00f6nelik, \u00f6\u011frenci taraf\u0131ndan olu\u015fturulan b\u00fcy\u00fck veri kaynaklar\u0131n\u0131 kullanmay\u0131 ama\u00e7layan ve giderek b\u00fcy\u00fcyen ara\u015ft\u0131rma alan\u0131na katk\u0131da bulunmaya \u00e7al\u0131\u015f\u0131yoruz. Bunu yapmak i\u00e7in, \u00f6ncelikle performans de\u011ferlendirme ve tahmin ile ilgili ara\u015ft\u0131rma hedeflerinden farkl\u0131 olarak, \u00f6\u011frencinin davran\u0131\u015fsal durumunun modellenmesine odaklanan bir ara\u015ft\u0131rma dizisi g\u00f6steriyoruz. Bir KA\u00c7D\u2019deki \u00f6\u011frencilerin ders videolar\u0131n\u0131 izlemek ya da forum yaz\u0131lar\u0131na cevap vermek ve bir sonraki eylemlerini tahmin etmek gibi t\u00fcm eylemlerini g\u00f6z \u00f6n\u00fcnde bulundurmak istiyoruz. B\u00f6yle bir yakla\u015f\u0131m, KA\u00c7D’larda toplanan ayr\u0131nt\u0131l\u0131, de\u011ferlendirme d\u0131\u015f\u0131 verileri kullan\u0131r ve seyir rehberli\u011fi arayan \u00f6\u011frenciler i\u00e7in bir tavsiye kayna\u011f\u0131 olarak hizmet etme potansiyeline sahiptir.<\/span><\/p>\n

                          KA\u00c7D’lere kat\u0131lan on binlerce \u00f6\u011frencinin t\u0131klama ak\u0131\u015f verilerini kullanarak, dersin sonunda ba\u015far\u0131l\u0131 olanlar\u0131n davran\u0131\u015flar\u0131n\u0131 modelleyerek, KA\u00c7D’ler aras\u0131nda gezinen \u00f6\u011frenciler aras\u0131nda genelle\u015ftirilebilir eylem \u00f6r\u00fcnt\u00fclerinin, ortaya \u00e7\u0131k\u0131p \u00e7\u0131kamayaca\u011f\u0131n\u0131 soruyoruz. Ba\u015far\u0131l\u0131 \u00f6\u011frencilerin KA\u00c7D’ler \u00fczerindeki trendlerini yakalamak, otomatik \u00f6neri sistemlerinin geli\u015ftirilmesini sa\u011flayabilir, b\u00f6ylece zorlanan \u00f6\u011frencilere ba\u015far\u0131l\u0131 olmak i\u00e7in harcad\u0131klar\u0131 zaman\u0131 optimize etmek i\u00e7in anlaml\u0131 ve etkili \u00f6neriler verilebilir. Bu g\u00f6rev i\u00e7in \u00fcretici s\u0131ral\u0131 modellerden yararlan\u0131r\u0131z. \u00dcretici dizilimli modeller girdi olarak bir olaylar diziliminde yer alabilir ve daha sonra ger\u00e7ekle\u015fecek olay \u00fczerinde olas\u0131l\u0131k da\u011f\u0131l\u0131m\u0131 olu\u015fturabilir. Bu \u00e7al\u0131\u015fmada di\u011fer \u00fcretici ve s\u0131ral\u0131 g\u00f6revlere uyguland\u0131\u011f\u0131nda geleneksel olarak ba\u015far\u0131l\u0131 olan \u00f6zellikle n-gram ve tekrarlayan sinir a\u011f\u0131 (TSA) modelleri olmak \u00fczere \u00fcretici s\u0131ral\u0131 modellerden iki t\u00fcr kullan\u0131lm\u0131\u015ft\u0131r.<\/span><\/p>\n

                          Bu b\u00f6l\u00fcm \u00f6zellikle, \u00f6\u011frencinin bir KA\u00c7D\u2019de ger\u00e7ekle\u015ftirdi\u011fi \u00f6nceki eylemler ba\u011flam\u0131nda verilen bir sonraki eylemi ne kadar iyi tahmin edebilece\u011fini analiz eder. Bu t\u00fcr bir analizin amac\u0131 nihayetinde, otomatik bir dan\u0131\u015fman\u0131n, \u00f6\u011frenciyi sonradan hangi eylemde bulunabilece\u011fi konusunda anlaml\u0131 bir rehberlik sa\u011flamak i\u00e7in modeli sorgulayabilece\u011fi bir sistem olu\u015fturmak olacakt\u0131r. Bir \u00e7ok durumda bir sonraki eylem,ders taraf\u0131ndan \u00f6ng\u00f6r\u00fclen bir sonraki kaynak olabilir ancak di\u011fer durumlarda, \u00f6\u011frencinin bilmedi\u011fi bir ders kitab\u0131n\u0131n bir k\u00f6\u015fesinde g\u00f6m\u00fcl\u00fc olan bir \u00f6nceki dersten veya zenginle\u015ftirme materyalinden bir kayna\u011fa ba\u015fvurmak bir tavsiye olabilir. E\u011fitti\u011fimiz bu bu modeller, \u00fcretici2<\/a><\/sup> olarak bilinirler, \u00e7\u00fcnk\u00fc \u00f6\u011frencinin daha \u00f6nce hangi eylemleri ger\u00e7ekle\u015ftirdi\u011fine ili\u015fkin \u00f6nceki bir ba\u011flam dikkate al\u0131nd\u0131\u011f\u0131nda hangi eylemin gelebilece\u011fini olu\u015fturmak i\u00e7in kullan\u0131labilirler. Eylemler, ders videosu a\u00e7ma, s\u0131nav sorusu cevaplama ya da bir forum g\u00f6nderisinde gezinme ve cevaplama gibi \u015feyleri i\u00e7erebilir. Bu ara\u015ft\u0131rma, KA\u00c7D’lerde potansiyel uygulamalarla ki\u015fiselle\u015ftirilmi\u015f dan\u0131\u015fmanlar olu\u015fturmaya y\u00f6nelik s\u0131ral\u0131, \u00fcretici modellerin birbirini takip eden verilerle di\u011fer e\u011fitim ba\u011flamlar\u0131na uygulanmas\u0131 i\u00e7in bir temel olarak hizmet vermektedir.<\/span><\/p>\n

                          \u0130LG\u0130L\u0130 \u00c7ALI\u015eMALAR<\/h2>\n

                          \u0130ngilizce dili s\u00f6z konusu oldu\u011funda, \u00fcretici modeller \u00f6rnek metinler olu\u015fturmak ve bu dilin nas\u0131l yap\u0131land\u0131r\u0131ld\u0131\u011f\u0131na dair \u00f6rnek metnin akla yatk\u0131nl\u0131\u011f\u0131n\u0131 de\u011ferlendirmek i\u00e7in kullan\u0131l\u0131r. Do\u011fal dil i\u015flemede kullan\u0131lan basit fakat g\u00fc\u00e7l\u00fc bir model (DD\u0130), olas\u0131l\u0131k da\u011f\u0131l\u0131m\u0131n\u0131n e\u011fitim k\u00fcmesinde her olas\u0131 n terim diziliminin \u00fczerine \u00f6\u011frenildi\u011fi bir n-gram modeldir (Brown, Desouza, Mercer, Pietra ve Lai, 1992). Son zamanlarda, tekrarlayan sinir a\u011flar\u0131 (TSA’lar), daha \u00f6nce g\u00f6r\u00fclen kelimelerin y\u00fcksek boyutlu s\u00fcrekli gizil bir duruma getirildi\u011fi bir sonraki kelime tahminini (Mikolov, Karafiat, Burget, Cernocky ve Khudanpur, 2010) ger\u00e7ekle\u015ftirmek i\u00e7in kullan\u0131lm\u0131\u015ft\u0131r. Bu gizli durum, daha \u00f6nce ba\u011flamda g\u00f6r\u00fclen kelimelerin hepsinin \u00f6zl\u00fc bir say\u0131sal temsilidir. Model daha sonra hangi kelimelerin geleceklerini tahmin etmek i\u00e7in bu g\u00f6sterimi kullanabilir. Bu \u00fcretici modellerin her ikisi de c\u00fcmleleri tamamlamak amac\u0131yla aday c\u00fcmleler ve kelimeler \u00fcretmek i\u00e7in kullan\u0131labilir. Bu \u00e7al\u0131\u015fmada, kelime ve c\u00fcmle dizilimlerinin akla yak\u0131nl\u0131k durumunu \u00f6\u011frenmek yerine, \u00fcretici modeller KA\u00c7D ba\u011flamlar\u0131nda \u00f6\u011frencilerin \u00fcstlendikleri eylem dizilimlerinin uygunlu\u011funu \u00f6\u011freneceklerdir. Daha sonra, bu t\u00fcr \u00fcretici modeller, \u00f6\u011frencinin daha sonra yapmas\u0131 gerekenler i\u00e7in tavsiyeler \u00fcretmek i\u00e7in kullan\u0131labilir.<\/span><\/p>\n

                          \u00d6\u011frenme analiti\u011fi toplulu\u011funda, \u00e7o\u011fu zaman KA\u00c7D ba\u011flamlar\u0131nda, \u00f6\u011frenciler taraf\u0131ndan \u00fcretilen verilerin analiz edildi\u011fi ilgili bir \u00e7al\u0131\u015fma vard\u0131r. Analitikler bir\u00e7ok farkl\u0131 \u00f6\u011frenci taraf\u0131ndan olu\u015fturulan veri t\u00fcr\u00fcyle ger\u00e7ekle\u015ftirilir ve pek \u00e7ok farkl\u0131 t\u00fcrde tahmin g\u00f6revi vard\u0131r. Crossley, Paquette, Dascalu, McNamara ve Baker (2016), bu durum i\u00e7in KA\u00c7D\u2019den al\u0131nan ham g\u00fcnl\u00fck kay\u0131tlar\u0131n\u0131n manuel \u00f6zellik m\u00fchendisli\u011fi s\u00fcreci ile \u00f6zetlendi\u011fi bir paradigma \u00f6rne\u011fi sunmaktad\u0131r. Bizim yakla\u015f\u0131m\u0131m\u0131zda, \u00f6zellik g\u00f6sterimleri do\u011frudan ham zaman serisi verilerinden \u00f6\u011frenilir. Bu yakla\u015f\u0131m, \u00f6zelliklerin geli\u015ftirilmesi i\u00e7in konu uzmanl\u0131\u011f\u0131 gerektirmez ve KA\u00c7D t\u0131klama ak\u0131\u015f\u0131ndaki ham bilgileri kullanmak i\u00e7in potansiyel olarak daha az kayb\u0131 olan bir yakla\u015f\u0131md\u0131r. Pardos ve Xu (2016), \u00f6nceki bilgilerin, KA\u00c7D kaynak kullan\u0131m\u0131 ile bilgi edinimi aras\u0131ndaki ili\u015fkinin geli\u015ftirilmesine yard\u0131mc\u0131 olmakta zorland\u0131\u011f\u0131n\u0131 belirledi. Bu \u00e7al\u0131\u015fmada, \u00f6\u011frencinin kendi kendini se\u00e7me durumu bir parazit ses ve ak\u0131l kar\u0131\u015ft\u0131r\u0131c\u0131l\u0131k kayna\u011f\u0131d\u0131r. Buna kar\u015f\u0131l\u0131k, \u00f6\u011frenen se\u00e7imi davran\u0131\u015fsal modellemede bir i\u015faret haline gelir. Reddy, Labutov ve Joachims (2016) ‘da, \u00e7evrimi\u00e7i bir ders sistemindeki \u00f6\u011frenci \u00f6\u011frenmesinin bir\u00e7ok y\u00f6n\u00fc, g\u00f6mme3<\/a><\/sup> yoluyla birlikte \u00f6zetlenmi\u015ftir. Bu g\u00f6mme i\u015flemi \u00f6devleri, \u00f6\u011frenci becerisini ve ders etkinli\u011fini d\u00fc\u015f\u00fck boyutlu bir uzaya e\u015fler. B\u00f6yle bir s\u00fcre\u00e7, modelin mevcut \u00f6\u011frenci yetene\u011fi tahminine dayanarak ders verme ve \u00f6dev verme yollar\u0131n\u0131n \u00f6nerilmesini sa\u011flar. Bu b\u00f6l\u00fcmdeki \u00e7al\u0131\u015fma ayn\u0131 zamanda \u00f6\u011frenciler i\u00e7in \u00f6\u011frenme yollar\u0131 \u00f6nermeyi ama\u00e7lamaktad\u0131r ancak forum sonras\u0131 eri\u015fimler ve ders video g\u00f6r\u00fcnt\u00fclemeleri gibi ek \u00f6\u011frenci davran\u0131\u015flar\u0131n\u0131n da modele d\u00e2hil edilmesi ile farkl\u0131la\u015f\u0131rlar. Ek olarak, farkl\u0131 \u00fcretici modeller kullan\u0131l\u0131r.<\/span><\/p>\n

                          Bu b\u00f6l\u00fcmde, yaln\u0131zca KA\u00c7D’lerden gelen kay\u0131t g\u00fcnl\u00fc\u011f\u00fc verileriyle \u00e7al\u0131\u015f\u0131yoruz. Bu kullan\u0131c\u0131 t\u0131klama ak\u0131\u015f\u0131 bir\u00e7ok etkile\u015fim alan\u0131n\u0131n \u00fcst\u00fcnden ge\u00e7erken, davran\u0131\u015f ara\u015ft\u0131rmas\u0131 \u00f6rnekleri bu etkile\u015fim dizilimlerine kat\u0131lan kaynaklar\u0131n i\u00e7eri\u011fini analiz etmi\u015ftir. Bu \u00f6rnekler, videonun etkile\u015fim d\u00fczeyini (Sharma, Biswas, Gandhi, Patil ve Deshmukh, 2016) karakterize etmek i\u00e7in KA\u00c7D video karelerinin analiz edilmesini, forum yaz\u0131lar\u0131n\u0131n i\u00e7eri\u011fini (Wen, Yang ve Rose, 2014; Reich, Stewart, Mavon ve Tingley, 2016) ve forumlardaki etkile\u015fimlerden kaynaklanan ve buna mahsus sosyal a\u011flar\u0131n analizini i\u00e7erir (Oleksandra ve Shane, 2016). T\u00fcm olas\u0131 \u00f6\u011frenci etkinlikleri kategorilerine bu i\u00e7erik odakl\u0131 yakla\u015f\u0131mlara k\u0131yasla daha soyut bir d\u00fczeyde bak\u0131yoruz.<\/span><\/p>\n

                          \u00d6\u011frenme analiti\u011fi ve EVM’deki bili\u015f a\u00e7\u0131s\u0131ndan, Bayesci bilgi takibi (BBT; Corbett ve Anderson, 1994) y\u00fczeysel ders yap\u0131s\u0131n\u0131 bir bilgi bile\u015fenleri kayna\u011f\u0131 olarak kullanarak modelin bir KA\u00c7D’ye iyile\u015ftirilerek uyarlanmas\u0131(Pardos, Bergner, Seaton ve Pritchard, 2013) gibi modeller arac\u0131l\u0131\u011f\u0131yla \u00f6\u011frencilerin mahrem bilgilerinin de\u011ferlendirilmesi i\u00e7in bir\u00e7ok \u00e7al\u0131\u015fma yap\u0131lm\u0131\u015ft\u0131r. Bu modelleme t\u00fcr\u00fc, \u00f6\u011frencilerin davran\u0131\u015flar\u0131n\u0131 \u00f6\u011frencilerin gizli bilgisini modellemek i\u00e7in \u00f6\u011frenme f\u0131rsatlar\u0131 olarak g\u00f6r\u00fcr. \u00c7al\u0131\u015fma ile ilgili olmas\u0131na ra\u011fmen, \u00f6\u011frenci bilgisi bu b\u00f6l\u00fcmde a\u00e7\u0131k\u00e7a modellenmemi\u015ftir. Bunun yerine, modellerimiz, \u00f6\u011frencinin davran\u0131\u015f verileri olan bu performans verilerinin tamamlay\u0131c\u0131s\u0131n\u0131 tahmin etmeye odaklan\u0131r.<\/span><\/p>\n

                          Derin bilgi takibi (DBT; Piech vd., 2015), tekrarlayan sinir a\u011flar\u0131n\u0131n, \u00f6\u011frencilerin \u00e7evrimi\u00e7i \u00f6\u011frenme ortamlar\u0131nda gezinirken daha \u00f6nce g\u00f6r\u00fclen de\u011ferlendirme sonu\u00e7lar\u0131na dayanarak s\u00fcrekli gizli bir temsilini olu\u015fturmak i\u00e7in kullan\u0131r. Bu \u00e7al\u0131\u015fmada, tekrarlayan sinir a\u011flar\u0131 karma\u015f\u0131k bir gizli durumu izleyerek bir \u00f6\u011frencinin \u00f6nceki de\u011ferlendirme sonu\u00e7lar\u0131n\u0131 \u00f6zetlemektedir. Bu \u00e7al\u0131\u015fma, y\u00fczeysel BBT yakla\u015f\u0131m\u0131na k\u0131yasla \u00f6\u011frenci bilgilerini temsil etmek i\u00e7in derinlemesine bir \u00f6\u011frenme yakla\u015f\u0131m\u0131n\u0131n kullan\u0131labilece\u011fini g\u00f6stermektedir. Bununla birlikte, bu sonu\u00e7lar\u0131n h\u00e2lihaz\u0131rda BBT’nin mevcut uzant\u0131lar\u0131yla a\u00e7\u0131klanaca\u011f\u0131 varsay\u0131lmaktad\u0131r (Khajah, Lindsey ve Mozer, 2016). Bilgiyi izlemeye yakla\u015fmada derin \u00f6\u011frenmenin kullan\u0131m\u0131 verilerde otomatik olarak hala yararl\u0131 ili\u015fkiler bulmaktad\u0131r ancak potansiyel olarak BBT i\u00e7in \u00f6nceden \u00f6nerilen uzant\u0131lara ili\u015fkin ek g\u00f6sterimler bulamam\u0131\u015ft\u0131r. Bu b\u00f6l\u00fcmdeki \u00e7al\u0131\u015fma, \u00f6\u011frencileri temsil etmek i\u00e7in derin a\u011flar\u0131n kullan\u0131lmas\u0131yla ilgilidir ancak yaln\u0131zca de\u011ferlendirme eylemlerinin kullan\u0131lmas\u0131 yerine, her t\u00fcrl\u00fc \u00f6\u011frenci eyleminin dikkate al\u0131nmas\u0131 bak\u0131m\u0131ndan farkl\u0131l\u0131k g\u00f6sterir.<\/span><\/p>\n

                          \u00d6zellikle, bu b\u00f6l\u00fcmde, n-gram yakla\u015f\u0131m\u0131n\u0131 ve uzun k\u0131sa vadeli haf\u0131za (UKVH) mimarisi olarak bilinen TSA varyant\u0131n\u0131 kullanmay\u0131 d\u00fc\u015f\u00fcnmekteyiz (Hochreiter ve Schmidhuber, 1997). Bu iki model hem veri dizilimini modeller ve hem de sonras\u0131nda hangi i\u015faretin gelmesi gerekti\u011fine dair bir olas\u0131l\u0131k da\u011f\u0131l\u0131m\u0131n\u0131 sa\u011flar. UKVH mimarilerinin ve benzer varyantlar\u0131n kullan\u0131m\u0131 k\u0131smen dizilimlerde uzun -ve k\u0131sa- aral\u0131klardaki ba\u011f\u0131ml\u0131l\u0131klar\u0131n yakalanmas\u0131na izin veren de\u011fi\u015ftirilebilir haf\u0131zas\u0131 sayesinde yak\u0131n zamanda, konu\u015fma, g\u00f6r\u00fcnt\u00fc ve metin analizi de d\u00e2hil olmak \u00fczere s\u0131ral\u0131 verileri i\u00e7eren \u00e7e\u015fitli alanlarda etkileyici sonu\u00e7lar elde etmi\u015ftir (Graves, Mohamed ve Hinton, 2013; Vinyals, Kaiser vd., 2015; Vinyals, Toshev, Bengio ve Erhan, 2015). \u00d6\u011frenci \u00f6\u011frenme davran\u0131\u015f\u0131, sabit bir eylem durum alan\u0131ndaki bir dizi eylem olarak temsil edilebildi\u011finden, ba\u015far\u0131l\u0131 \u00f6\u011frenmeyi karakterize eden karma\u015f\u0131k \u00f6r\u00fcnt\u00fcleri yakalamak i\u00e7in UKVH’ler potansiyel olarak kullan\u0131labilir. \u00d6nceki \u00e7al\u0131\u015fmalarda, \u00f6\u011frenci t\u0131klama verilerinin modellenmesi n-gram modelleri gibi y\u00f6ntemlerle \u00fcmit verici olmu\u015ftur(Wen ve Rose, 2014).<\/span><\/p>\n

                          VER\u0130 K\u00dcMES\u0130<\/h2>\n

                          Bu b\u00f6l\u00fcmde kullan\u0131lan veri k\u00fcmesi, 2013 Bahar d\u00f6nemi \u0130statistik BerkeleyX KA\u00c7D’den geldi. KA\u00c7D be\u015f hafta boyunca video konferanslar\u0131, ev \u00f6devleri, tart\u0131\u015fma forumlar\u0131 ve iki s\u0131navla devam etti. Orijinal veri k\u00fcmesi, her \u00f6\u011frencinin bir \u015fekilde KA\u00c7D ile etkile\u015fime giren bir kullan\u0131c\u0131 kayd\u0131 oldu\u011fu, 31.000 \u00f6\u011frenciden gelen 17 milyon olay\u0131 i\u00e7ermektedir. Bu etkile\u015fimler, derste belirli bir URL’ye gitme, bir forum mesajlar\u0131nda oylamaya kat\u0131lma, bir s\u0131nav sorusunu cevaplama ve bir konferans videosu oynatma gibi olaylar\u0131 i\u00e7erir. Veriler, her bir kullan\u0131c\u0131n\u0131n t\u00fcm olaylar\u0131n\u0131n s\u0131ralamal\u0131 olarak toplanabilmi\u015f olmas\u0131 i\u00e7in i\u015flenir: 3687 olay t\u00fcr\u00fc m\u00fcmk\u00fcnd\u00fcr. Veri k\u00fcmesindeki her sat\u0131r, ger\u00e7ekle\u015ftirilen eylemi veya \u00f6\u011frenci taraf\u0131ndan eri\u015filen URL’yi temsil eden belirli bir dizine d\u00f6n\u00fc\u015ft\u00fcr\u00fcl\u00fcr.<\/span><\/p>\n

                          Bu nedenle, her bir kullan\u0131c\u0131n\u0131n i\u015flem grubu, 3687 \u00f6zg\u00fcn t\u00fcr olan bir dizilim indisle temsil edilir. Kay\u0131tl\u0131 etkinlik ge\u00e7mi\u015fimiz, forumun ba\u015fl\u0131klar\u0131, s\u0131navlar, video sayfalar\u0131 ve wiki sayfalar\u0131 gibi dersin farkl\u0131 sayfalar\u0131na giden \u00f6\u011frencileri i\u00e7eriyordu. Bu sayfalarda, video oynatmak ve duraklatmak ya da bir sorunu kontrol etmek gibi sayfa i\u00e7inde yap\u0131lan eylemleri de kaydettik. Ayr\u0131ca s\u0131ral\u0131 olaylar ad\u0131 verilen JavaScript gezinmelerini de kaydederiz. \u00d6n i\u015fleme s\u00fcrecimizin bu a\u00e7\u0131klamas\u0131nda, bu olay dizilimlerini, art arda gelen olay taraf\u0131ndan y\u00f6nlendirilen URL ile a\u00e7\u0131k\u00e7a ili\u015fkilendirilmeden, kendi ba\u015flar\u0131na kaydederiz. Tablo 19.1, veri k\u00fcmesinde mevcut olan farkl\u0131 olay t\u00fcrlerini ve olaya ba\u011fl\u0131 belirli bir URL’yi ili\u015fkilendirmeyi se\u00e7ip se\u00e7medi\u011fimizi listelemektedir. \u00d6n i\u015flemimizde, bu olaylar\u0131n baz\u0131lar\u0131 URL’ye \u00f6zg\u00fc olarak kaydedilir, bu modelin \u00f6\u011frencinin bu olaylar i\u00e7in eri\u015fti\u011fi tam URL\u2019ye maruz kalaca\u011f\u0131 anlam\u0131na gelir. Baz\u0131 olaylar URL’ye \u00f6zg\u00fc olmayan olarak kaydedilir; bu, modelin yaln\u0131zca eylemin ger\u00e7ekle\u015fti\u011fini bildi\u011fi ancak o eylemin derste hangi URL’ye ba\u011fl\u0131 oldu\u011funu bilmedi\u011fi anlam\u0131na gelir. Orijinal veri k\u00fcmesinde 40 kattan daha az ger\u00e7ekle\u015fen olaylar\u0131n ayr\u0131 tutuldu\u011funu da dikkate al\u0131n\u0131z. Bu nedenle, forum etkinliklerinin bir\u00e7o\u011fu URL’ye \u00f6zg\u00fc olduklar\u0131 ancak \u00e7ok s\u0131k ger\u00e7ekle\u015fmedikleri i\u00e7in ayr\u0131 tutulmu\u015flard\u0131r. Git dizilimi, sonraki dizilimi ve \u00f6nceki dizilimi, \u00f6\u011frenciler taray\u0131c\u0131 sayfas\u0131nda g\u00f6r\u00fcnen gezinme d\u00fc\u011fmelerini se\u00e7tiklerinde tetiklenen olaylara at\u0131fta bulunur. Sonraki dizilimi ve \u00f6nceki dizilimi, s\u0131ras\u0131yla dersteki \u00f6nceki veya sonraki i\u00e7erik sayfas\u0131na gider. Git dizilimi, bir b\u00f6l\u00fcm i\u00e7inde bir alt b\u00f6l\u00fcm i\u00e7inden ba\u015fka bir alt b\u00f6l\u00fcme atlamay\u0131 temsil eder.<\/span><\/p>\n

                          <\/a><\/a><\/a><\/a><\/a> Tablo 19.1. G\u00fcnl\u00fc\u011fe Kaydedilen Olay Tipleri ve \u00d6zellikleri<\/i><\/span><\/span><\/p>\n\n\n<\/colgroup>\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
                          Ders Sayfas\u0131 Olaylar\u0131<\/b><\/span><\/span><\/td>\n<\/tr>\n
                          Sayfa G\u00f6r\u00fcn\u00fcm\u00fc (URL’ye \u00f6zg\u00fc) <\/span><\/span><\/td>\n<\/tr>\n
                          Git dizilimi (URL’ye \u00f6zg\u00fc de\u011fil) <\/span><\/span><\/td>\n<\/tr>\n
                          Sonraki dizilimi (URL ye \u00f6zg\u00fc de\u011fil) <\/span><\/span><\/td>\n<\/tr>\n
                          Dizilim \u00d6nceki (URL’ ye \u00f6zg\u00fc de\u011fil)<\/span><\/span><\/td>\n<\/tr>\n
                          Wiki Olaylar\u0131 <\/b><\/span><\/span><\/td>\n<\/tr>\n
                          Sayfa G\u00f6r\u00fcn\u00fcm\u00fc (URL’ye \u00f6zg\u00fc) <\/span><\/span><\/td>\n<\/tr>\n
                          Video Olaylar\u0131<\/b><\/span><\/span><\/td>\n<\/tr>\n
                          Video Duraklat (URL’ye \u00f6zg\u00fc de\u011fil) <\/span><\/span><\/td>\n<\/tr>\n
                          Video Oynat\u0131m\u0131 (URL’ye \u00f6zg\u00fc de\u011fil) <\/span><\/span><\/td>\n<\/tr>\n
                          Problem Olaylar\u0131 <\/b><\/span><\/span><\/td>\n<\/tr>\n
                          Sayfa G\u00f6r\u00fcn\u00fcm\u00fc (URL’ye \u00f6zg\u00fc de\u011fil) <\/span><\/span><\/td>\n<\/tr>\n
                          Sorun Kontrolu (URL’ye \u00f6zg\u00fc de\u011fil) <\/span><\/span><\/td>\n<\/tr>\n
                          Problem Cevap G\u00f6ster(URL’ye \u00f6zg\u00fc de\u011fil)<\/span><\/span><\/td>\n<\/tr>\n
                          Forum Olaylar\u0131<\/b><\/span><\/span><\/td>\n<\/tr>\n
                          Forum G\u00f6r\u00fcn\u00fcm\u00fc (URL’ye \u00d6zg\u00fc)<\/span><\/span><\/td>\n<\/tr>\n
                          Forum Kapat (filtrelenerek ay\u0131r\u0131ld\u0131)<\/span><\/span><\/td>\n<\/tr>\n
                          Forum Olu\u015ftur (filtrelenerek ayr\u0131ld\u0131 )<\/span><\/span><\/td>\n<\/tr>\n
                          Forum Silme (filtrelenerek ayr\u0131ld\u0131<\/span><\/span><\/td>\n<\/tr>\n
                          Forum Teyit (filtrelenerek ayr\u0131ld\u0131)<\/span><\/span><\/td>\n<\/tr>\n
                          Forum \u0130zlemesi (URL’ye \u00d6zg\u00fc)<\/span><\/span><\/td>\n<\/tr>\n
                          Forum Yan\u0131t\u0131 (URL’ye \u00d6zg\u00fc)<\/span><\/span><\/td>\n<\/tr>\n
                          Forum Aramas\u0131 (URL\u2019ye \u00f6zg\u00fc olmayan)<\/span><\/span><\/td>\n<\/tr>\n
                          Forum Takip Etme (filtrelenerek ayr\u0131ld\u0131)<\/span><\/span><\/td>\n<\/tr>\n
                          Forum Oy Kullanmama (filtrelenerek ayr\u0131ld\u0131)<\/span><\/span><\/td>\n<\/tr>\n
                          Forum G\u00fcncellemesi (filtrelenerek ayr\u0131ld\u0131)<\/span><\/span><\/td>\n<\/tr>\n
                          Forum Be\u011feni (URL’ye \u00d6zg\u00fc)<\/span><\/span><\/td>\n<\/tr>\n
                          Forum \u0130zlenen Konular\u0131 G\u00f6r\u00fcnt\u00fcle (URL’ye \u00d6zel)<\/span><\/span><\/td>\n<\/tr>\n
                          Forum Sat\u0131r \u0130\u00e7i G\u00f6r\u00fcnt\u00fcle (URL’ye \u00d6zel)<\/span><\/span><\/td>\n<\/tr>\n
                          Forum Kullan\u0131c\u0131 Profili G\u00f6r\u00fcnt\u00fcle (URL’ye \u00d6zel)<\/span><\/span><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

                          \u00d6rne\u011fin, bir \u00f6\u011frenci 2. b\u00f6l\u00fcm 1. alt b\u00f6l\u00fcm URL\u2019sine eri\u015firse, bir ders videosu oynat\u0131rsa, bir sonraki ok d\u00fc\u011fmesine (bir sonraki alt b\u00f6l\u00fcme eri\u015fmek i\u00e7in bir JavaScript gezinmesi ger\u00e7ekle\u015ftirir) t\u0131klar, bir s\u0131nav sorusunu cevaplar ve ard\u0131ndan gezinti \u00e7ubu\u011funda (ba\u015fka bir JavaScript gezintisi ger\u00e7ekle\u015ftiren) 5. alt b\u00f6l\u00fcme t\u0131klar, bu \u00f6\u011frencinin dizilimi be\u015f farkl\u0131 indisle temsil edilir. Birincisi b\u00f6l\u00fcm 2 alt b\u00f6l\u00fcm 1’in URL’sine, ikincisi bir oynatma videosu belirtecine, \u00fc\u00e7\u00fcnc\u00fcs\u00fc bir sonraki gezinti olay\u0131na, d\u00f6rd\u00fcnc\u00fcs\u00fc \u00f6\u011frencinin kursa d\u00e2hil oldu\u011fu belirli probleme gitme olay\u0131 be\u015fincisi gezinti olay\u0131 gite tekab\u00fcl eder. Modele s\u0131rayla bu be\u015f indisin bir listesi verilecek ve sonra neyin gelmesi gerekti\u011fini tahmin etmek i\u00e7in e\u011fitilecektir. Bu nedenle indisler, \u00f6\u011frencinin ger\u00e7ekle\u015ftirdi\u011fi eylem dizilimini temsil eder. Uzunluk i\u00e7in be\u015f gerekli de\u011fildir; \u00fcretici modellere iste\u011fe ba\u011fl\u0131 uzunluktaki dizilimler verilebilir.<\/span><\/p>\n

                          31.000 \u00f6\u011frenciden, 8094’\u00fc dersin \u00f6\u011fretenleri taraf\u0131ndan \u201conayl\u0131\u201d kabul edilebilecek miktarda \u00f6devi tamamlad\u0131 ve s\u0131navlarda yeterince y\u00fcksek puan ald\u0131lar. Di\u011fer KA\u00c7D ba\u011flamlar\u0131nda, sertifikan\u0131n bazen \u00f6\u011frencinin \u00f6zel bir sertifika i\u00e7in para \u00f6demesi anlam\u0131na geldi\u011fini ancak bu KA\u00c7D i\u00e7in ge\u00e7erli olmad\u0131\u011f\u0131n\u0131 unutmay\u0131n\u0131z. Sertifikal\u0131 \u00f6\u011frenciler, orijinal 17 milyon etkinli\u011fin 11.2 milyonunu olu\u015ftururken, sertifikal\u0131 \u00f6\u011frenci ba\u015f\u0131na ortalama 1390 etkinlik ger\u00e7ekle\u015fti. Sertifikal\u0131 ve sertifikal\u0131 olmayanlar aras\u0131ndaki ayr\u0131m, bu model i\u00e7in \u00f6nemlidir, \u00e7\u00fcnk\u00fc sertifikal\u0131 olarak kabul edilen \u00f6\u011frencilerin eylemlerinin bu KA\u00c7D i\u00e7in makul bir \u015fekilde ba\u015far\u0131l\u0131 bir navigasyon bilgisi \u00f6r\u00fcnt\u00fcs\u00fc olabilece\u011fi hipotezine g\u00f6re \u00fcretici modelleri yaln\u0131zca “sertifikal\u0131” kabul edilen \u00f6\u011frencilerin ger\u00e7ekle\u015ftirdi\u011fi eylem dizilimleri \u00fczerine e\u011fitmeyi se\u00e7tik.<\/span><\/p>\n

                          Veri k\u00fcmesindeki her sat\u0131r, kullan\u0131c\u0131n\u0131n eri\u015fiminin tam URL’si, kullan\u0131c\u0131 i\u00e7in benzersiz bir tan\u0131mlay\u0131c\u0131, i\u015flemin tam olarak ger\u00e7ekle\u015fti\u011fi zaman ve daha fazlas\u0131 gibi i\u015flemle ilgili bilgileri i\u00e7erir. Bu b\u00f6l\u00fcm i\u00e7in, zaman veya muhtemel di\u011fer ba\u011flamsal bilgileri dikkate almay\u0131z, bunun yerine sadece \u00f6\u011frencinin eri\u015fti\u011fi kayna\u011fa veya \u00f6\u011frencinin yapt\u0131\u011f\u0131 eyleme odaklan\u0131r\u0131z. Veri k\u00fcmesinin tamam\u0131nda 40 kattan daha az ger\u00e7ekle\u015fen olaylar, nadiren eri\u015filen tart\u0131\u015fma g\u00f6nderileri veya kullan\u0131c\u0131 profili ziyaretleri olduklar\u0131ndan ve KA\u00c7D’de gezinen di\u011fer \u00f6\u011frencilere uygulanmalar\u0131 muhtemel olmad\u0131\u011f\u0131ndan kald\u0131r\u0131ld\u0131.<\/span><\/p>\n

                          Y\u00d6NTEM<\/h2>\n

                          Bu \u00e7al\u0131\u015fmada, iki \u00fcretici modelin, tekrarlayan sinir a\u011f\u0131 mimarisi ve n-gram kullan\u0131m\u0131n\u0131 ara\u015ft\u0131rd\u0131k. Bu b\u00f6l\u00fcmde, tekrarlayan sinir a\u011f\u0131n\u0131n mimarisini ve UKVH uzant\u0131s\u0131n\u0131 ayr\u0131nt\u0131lar\u0131yla a\u00e7\u0131kl\u0131yoruz, hipotez olarak sundu\u011fumuz model bir sonraki eylem tahmininde en iyi performans\u0131 g\u00f6sterecektir. N-gram gibi di\u011fer \u201cy\u00fczeysel\u201d modeller daha sonra a\u00e7\u0131klanmaktad\u0131r.<\/span><\/p>\n

                          Tekrarlayan Sinir A\u011flar\u0131<\/h3>\n

                          Tekrarlayan sinir a\u011flar\u0131 (TSA’lar), keyfi uzunluktaki s\u0131ral\u0131 verileri i\u015flemek i\u00e7in tasarlanm\u0131\u015f bir sinir a\u011f\u0131 modelleri ailesidir. Tekrarlayan sinir a\u011flar\u0131, belirli bir diziliminin i\u015flenmesi s\u00fcresince devam eden s\u00fcrekli ve gizli bir durumu etraf\u0131nda tutarak \u00e7al\u0131\u015f\u0131r. Bu gizli durum, \u015fimdiye kadarki dizilim ile ilgili bilgileri yakalar, b\u00f6ylece dizimin sonraki b\u00f6l\u00fcmlerindeki \u00f6ng\u00f6r\u00fc, bu s\u00fcrekli gizli durumdan etkilenebilir. Ad\u0131ndan da anla\u015f\u0131laca\u011f\u0131 gibi, TSA’lar ileri beslemeli sinir a\u011flar\u0131 taraf\u0131ndan kullan\u0131lan hesaplama yakla\u015f\u0131m\u0131n\u0131 kullan\u0131r ve ayn\u0131 zamanda zaman ad\u0131mlar\u0131 aras\u0131nda devam eden s\u00fcrekli bir gizli durumu da dayat\u0131r. Gizli durumu bir giri\u015f dizilimindeki elemanlar aras\u0131nda tutmak, tekrarlayan sinir a\u011flar\u0131na s\u0131ral\u0131 modelleme g\u00fcc\u00fc veren \u015feydir. Bu \u00e7al\u0131\u015fmada, TSA’ya her girdi KA\u00c7D veri k\u00fcmesinden gelen gran\u00fcl bir \u00f6\u011frenci olay\u0131 olacakt\u0131r. TSA, \u015fimdiye kadar g\u00f6r\u00fclen olaylara dayanarak \u00f6\u011frencilerin bir sonraki olay\u0131n\u0131 tahmin etmek i\u00e7in e\u011fitilmi\u015ftir. \u015eekil 19.1, girdilerin \u00f6\u011frencilerin eylemleri olaca\u011f\u0131 ve \u00e7\u0131kt\u0131lar\u0131n dizilim i\u00e7inden bir sonraki \u00f6\u011frenci hareketi olaca\u011f\u0131 basit bir TSA diyagram\u0131n\u0131 g\u00f6sterir. A\u015fa\u011f\u0131daki denklemler, TSA modelinin parametrelerinin her birinde kullan\u0131lan matematiksel i\u015flemleri g\u00f6sterir: ht, s\u00fcrekli gizli durumu temsil eder. Bu gizli durum, xt + 1’deki tahminin gizli durum ht’den etkilenece\u011fi \u015feklinde bulundurulur. TSA modeli, bir giri\u015f a\u011f\u0131rl\u0131\u011f\u0131 matrisi Wx, tekrarlayan a\u011f\u0131rl\u0131k matrisi Wh, ba\u015flang\u0131\u00e7 durumu h0 ve \u00e7\u0131k\u0131\u015f matrisi Wy ile parametrelendirilir: bh ile by s\u0131ras\u0131yla gizli ve \u00e7\u0131k\u0131\u015f birimleri i\u00e7in sapmalard\u0131r.<\/span><\/p>\n

                          ht<\/sub> = tanh(W x<\/sub>xt<\/sub> + W h<\/sub>ht\u22121<\/sub> + bh<\/sub>) (1) <\/span><\/p>\n

                          yt<\/sub> =\u03c3(W y<\/sub>ht<\/sub> +by<\/sub>) (2)<\/span><\/p>\n

                          \"\"<\/p>\n

                          <\/a> \u015eekil 19.1. Basit tekrarlayan sinir a\u011f\u0131<\/i><\/span><\/span><\/p>\n

                          UKVH Modelleri<\/h3>\n

                          TSA’n\u0131n pop\u00fcler bir \u00e7e\u015fidi, gizli durumda ne zaman anlaml\u0131 bilgiler tutaca\u011f\u0131n\u0131 ve gizli durumu ne zaman temizleyece\u011fi veya ne zaman “unutaca\u011f\u0131n\u0131” \u00f6\u011frenen \u201ckap\u0131lar\u201d ekleyerek TSA\u2019lar\u0131n uzun d\u00f6nem ba\u011f\u0131ml\u0131l\u0131klar\u0131 \u00f6\u011frenmelerine yard\u0131mc\u0131 oldu\u011fu d\u00fc\u015f\u00fcn\u00fclen, anlaml\u0131 uzun vadeli etkile\u015fimlerin s\u00fcreklili\u011fine izin veren uzun k\u0131sa s\u00fcreli bellek (UKVH; Hochreiter ve Schmidhuber, 1997) mimarisidir. UKVH’ler, gizli durumun ne zaman temizlenece\u011fini ve ne zaman yararl\u0131 bilgilerle g\u00fc\u00e7lenece\u011fini belirlemek i\u00e7in a\u00e7\u0131k\u00e7a \u00f6\u011frenilen ek ge\u00e7it parametreleri ekler. Bunun yerine, her gizli durum, h1<\/sub> ek ge\u00e7it parametreleri i\u00e7eren bir UKVH h\u00fccre birimi ile de\u011fi\u015ftirilir. Bu kap\u0131lar nedeniyle, UKVH’lerin basit TSA’lardan daha etkili bir \u015fekilde e\u011fitildi\u011fi bulunmu\u015ftur (Bengio, Simard ve Frasconi, 1994; Gers, Schmidhuber ve Cummins, 2000). Bir UKVH i\u00e7in g\u00fcncelleme denklemleri a\u015fa\u011f\u0131daki gibidir:<\/span><\/p>\n

                          f<\/i><\/span>t<\/span><\/sub> = <\/span>\u03c3<\/span>(<\/span>W <\/i><\/span>fx<\/span><\/sub>x<\/i><\/span>t <\/span><\/sub>+ <\/span>W <\/i><\/span>fh<\/span><\/sub>h<\/i><\/span>t <\/span><\/sub>\u2212 1<\/i><\/span><\/sub> <\/i>+ <\/span>b<\/i><\/span>f<\/span><\/sub>) (3) <\/span><\/span><\/span><\/p>\n

                          i<\/i><\/span>t<\/span><\/sub> = \u03c3(<\/span>W<\/i><\/span>ix<\/span><\/sub>x<\/i><\/span>t<\/span><\/sub> <\/i>+ <\/span>W<\/i><\/span>ih<\/span><\/sub>h<\/i><\/span>t<\/span><\/sub>\u22121 <\/i><\/span><\/sub>+ <\/span>b<\/i><\/span>i<\/span><\/sub>) (4) <\/span><\/span><\/span><\/p>\n

                          C<\/i><\/span>~<\/i><\/span><\/sup>t<\/i><\/span><\/sub> <\/i>= <\/span>tanh<\/i><\/span>(<\/span>W<\/i><\/span>Cx<\/i><\/span><\/sub>x<\/i><\/span>t<\/i><\/span><\/sub> <\/i>+ <\/span>W <\/i><\/span>Ch<\/i><\/span><\/sub>h<\/i><\/span>t\u22121<\/i><\/span><\/sub> +<\/span>b<\/i><\/span>C<\/i><\/span><\/sub>) <\/span>(5)<\/span> <\/span><\/span><\/p>\n

                          Ct <\/i><\/span>= <\/span>ft <\/i><\/span>\u00d7 <\/span>Ct<\/i><\/span>\u22121 + <\/span>it <\/i><\/span>\u00d7 <\/span>C\u02dc<\/i><\/span>t<\/i><\/span><\/sub> <\/i>(6) <\/span><\/span><\/span><\/p>\n

                          ot <\/i><\/span>= <\/span>\u03c3<\/span> (<\/span>W <\/i><\/span>ox<\/i><\/span><\/sub>x<\/i><\/span>t<\/i><\/span><\/sub> <\/i>+ <\/span>W <\/i><\/span>oh<\/i><\/span><\/sub>h<\/i><\/span>t<\/i><\/span><\/sub>\u22121 <\/span><\/sub>+ <\/span>b<\/i><\/span>o<\/i><\/span><\/sub>) <\/span>(7) <\/span><\/span><\/span><\/p>\n

                          ht <\/i><\/span>= <\/span>o<\/i><\/span>t<\/i><\/span><\/sub> <\/i>\u00d7 <\/span>tanh<\/i><\/span>(<\/span>Ct<\/i><\/span>) (8)<\/span><\/span><\/span><\/p>\n

                          \"\"<\/p>\n

                          <\/a> \u015eekil 19.2. UKVH i\u00e7in g\u00fcncelleme denklemlerine kar\u015f\u0131l\u0131k gelen say\u0131larla bir h\u00fccrenin anatomisi.<\/i><\/span><\/span><\/p>\n

                          \u015eekil 19.2, \u015fekildeki say\u0131lar\u0131n UKVH: ft, it ve ot i\u00e7in, daha \u00f6nce bahsedilen g\u00fcncelleme denklemlerine kar\u015f\u0131l\u0131k geldi\u011fi bir h\u00fccrenin anatomisini g\u00f6sterir; bu, UKVH taraf\u0131ndan \u00f6nceki h\u00fccreden “unutma” verilerini ve yeni h\u00fccre durumuna neyin “girilece\u011fini” ve h\u00fccre durumundan neyin “\u00e7\u0131kaca\u011f\u0131n\u0131” belirlemek i\u00e7in kullan\u0131lan ge\u00e7it mekanizmalar\u0131n\u0131 temsil eder. Ct, bilgilerin UKVH’yi beslemesi s\u0131ras\u0131nda bilgilerin \u00e7\u0131kar\u0131ld\u0131\u011f\u0131 ve eklendi\u011fi gizli h\u00fccre durumunu temsil eder. C\u02dct, bir sonraki h\u00fccre durumunu g\u00fcncellemek i\u00e7in ge\u00e7itli hale getirilmi\u015f olan yeni aday h\u00fccre durumunu temsil eder.<\/span><\/p>\n

                          UKVH Uygulamas\u0131<\/h3>\n

                          Bu b\u00f6l\u00fcmde kullan\u0131lan \u00fcretici UKVH modelleri, Theano’nun \u00fczerine in\u015fa edilmi\u015f bir Python k\u00fct\u00fcphanesi olan Keras (Chollet, 2015) kullan\u0131larak ger\u00e7ekle\u015ftirilmi\u015ftir (Bergstra vd., 2010; Bastien vd., 2012). Model, bir indeks numaras\u0131 ile temsil edilen her \u00f6\u011frenci eylemini ger\u00e7ekle\u015ftirir. Bu indisler, yapay de\u011fi\u015fkenlik olarak da bilinen 1-s\u0131cak vekt\u00f6r kodlamas\u0131ndaki indekse kar\u015f\u0131l\u0131k gelir. Model her dizini bir g\u00f6mme vekt\u00f6r\u00fcne d\u00f6n\u00fc\u015ft\u00fcr\u00fcr ve sonra g\u00f6m\u00fcl\u00fc vekt\u00f6r\u00fc birer birer t\u00fcketir. G\u00f6mme katman\u0131n\u0131n kullan\u0131m\u0131 s\u00f6zc\u00fckleri \u00e7ok boyutlu semantik bir uzaya e\u015flemenin bir yolu olarak do\u011fal dil i\u015fleme s\u00fcre\u00e7lerinde ve dil modellemede yayg\u0131nd\u0131r (Goldberg ve Levy, 2014). Burada bir g\u00f6mme katman\u0131, KA\u00c7D eylem alan\u0131ndaki eylemler i\u00e7in benzer bir e\u015flemenin olabilece\u011fi hipotezi ile birlikte kullan\u0131lmaktad\u0131r. Model, daha \u00f6nce \u00f6\u011frenci taraf\u0131ndan ger\u00e7ekle\u015ftirilen eylemler g\u00f6z \u00f6n\u00fcne al\u0131nd\u0131\u011f\u0131nda, bir sonraki \u00f6\u011frenci eylemini tahmin etmek i\u00e7in e\u011fitilmi\u015ftir. Zaman i\u00e7inde geri yay\u0131l\u0131m (Werbos, 1988), UKVH parametrelerini e\u011fitmek i\u00e7in kullan\u0131l\u0131r, bir sonraki eylemin indeksi olan bir ger\u00e7ek referans de\u011fer olarak softmax katman\u0131 kullan\u0131l\u0131r. Kay\u0131p hesaplan\u0131rken kategorik \u00e7apraz entropi, RMSprop ise en iyile\u015ftirici olarak kullan\u0131l\u0131r. UKVH katmanlar\u0131 aras\u0131na a\u015f\u0131r\u0131 uyumu engelleme y\u00f6ntemi olarak b\u0131rakma katmanlar\u0131 eklenmi\u015ftir (Pham, Bluche, Kermorvant ve Louradour, 2014). Her bir e\u011fitim verisi grubu i\u00e7in rastgele s\u0131f\u0131rlama y\u00fczdesi a\u011f kenar\u0131 a\u011f\u0131rl\u0131klar\u0131n\u0131n belirli bir y\u00fczdelik grubu s\u0131f\u0131rlar. Gelecekteki \u00e7al\u0131\u015fmalarda, \u00f6zellikle UKVH’ler ve TSA’lar i\u00e7in haz\u0131rlanm\u0131\u015f di\u011fer d\u00fczenlile\u015ftirme tekniklerini de\u011ferlendirmek faydal\u0131 olabilir (Zaremba, Sutskever ve Vinyals, 2014). Yaln\u0131zca veri dizisindeki gezinme eylemlerini \u00e7\u0131karmakla ba\u015flayan \u00f6n i\u015fleme ve UKVH model kodumuzun bir versiyonunu kamuya a\u00e7\u0131k hale getirdik4<\/a><\/sup>.<\/span><\/p>\n

                          UKVH Hiperparametre Aramas\u0131<\/h3>\n

                          \u0130lk ara\u015ft\u0131rmam\u0131z\u0131n bir par\u00e7as\u0131 olarak 24 UKVH modelinin her birini farkl\u0131 hiperparametre k\u00fcmeleri ile 10 evrede5<\/a><\/sup> e\u011fittik. Bir evre, veriler aras\u0131nda tam bir ge\u00e7i\u015f yapan parametre yerle\u015ftirme algoritmas\u0131d\u0131r. UKVH modellerimiz i\u00e7in aranan hiperparametre alan\u0131 Tablo 19.2’de g\u00f6sterilmektedir. Bu hiperparametreler, etki b\u00fcy\u00fckl\u00fc\u011f\u00fcne g\u00f6re farkl\u0131 hiperparametrelere \u00f6ncelik veren \u00f6nceki \u00e7al\u0131\u015fmalara dayanarak \u015febeke aramalar\u0131 i\u00e7in se\u00e7ilmi\u015ftir (Greff, Srivastava, Koutnik, Steunebrink ve Schmidhuber, 2015). Zamanlama ad\u0131na, 3 katl\u0131 UKVH modellerini.0001 \u00f6\u011frenme oranlar\u0131 ile e\u011fitmemeyi tercih ettik. Ayr\u0131ca, ek hiperparametreyi ve e\u011fitim y\u00f6ntemlerini ara\u015ft\u0131rmak i\u00e7in ba\u015flang\u0131\u00e7 noktas\u0131 olarak hizmet etmek \u00fczere ilk incelemenin sonu\u00e7lar\u0131n\u0131 kulland\u0131\u011f\u0131m\u0131z geni\u015fletilmi\u015f bir ara\u015ft\u0131rma yapt\u0131k.<\/span><\/p>\n

                          E\u011fitim TSA’lar\u0131 nispeten zaman al\u0131c\u0131 oldu\u011fundan, geni\u015fletilmi\u015f ara\u015ft\u0131rma, \u00fcmit vaat eden hiperparametre kombinasyonlar\u0131n\u0131n bir alt k\u00fcmesinden olu\u015fuyordu (Sonu\u00e7lar b\u00f6l\u00fcm\u00fcne bak\u0131n).<\/span><\/p>\n

                            \n
                          1. \n
                              \n
                            1. \n
                                \n
                              1. \n
                                  \n
                                1. \n

                                  <\/a><\/a><\/a><\/a><\/a> Tablo 19.2. UKVH Hiperparametre G\u00f6zene\u011fi<\/i><\/span><\/span><\/p>\n<\/li>\n<\/ol>\n<\/li>\n<\/ol>\n<\/li>\n<\/ol>\n<\/li>\n<\/ol>\n\n\n\n\n\n<\/colgroup>\n\n\n\n\n
                                  \n

                                  Gizli Katmanlar<\/b><\/span><\/p>\n<\/td>\n

                                  		\n

                                  1 <\/b><\/span><\/p>\n<\/td>\n

                                  		\n

                                  2 <\/b><\/span><\/p>\n<\/td>\n

                                  		\n

                                  3<\/b><\/span><\/p>\n<\/td>\n<\/tr>\n

                                  \n

                                  Gizli Katmandaki D\u00fc\u011f\u00fcmler<\/b><\/span><\/p>\n<\/td>\n

                                  		\n

                                  64 <\/span><\/p>\n<\/td>\n

                                  		\n

                                  128 <\/span><\/p>\n<\/td>\n

                                  		\n

                                  256<\/span><\/p>\n<\/td>\n<\/tr>\n

                                  \n

                                  \u00d6\u011frenme oran\u0131 (_)<\/b><\/span><\/p>\n<\/td>\n

                                  		\n

                                  .001<\/span><\/p>\n<\/td>\n

                                  		\n

                                  .0001 <\/span><\/p>\n<\/td>\n

                                  		\n

                                  .0001*<\/span><\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

                                  \u00c7apraz Do\u011frulama<\/h3>\n

                                  Her modelin yorday\u0131c\u0131 g\u00fcc\u00fcn\u00fc de\u011ferlendirmek i\u00e7in 5 katl\u0131 \u00e7apraz do\u011frulama kullan\u0131lm\u0131\u015ft\u0131r. Her model verinin %80’i i\u00e7in e\u011fitildi ve sonra kalan %20’si \u00fczerinde do\u011fruland\u0131; Bu be\u015f kez yap\u0131ld\u0131, b\u00f6ylece her \u00f6\u011frenci eylemi bir kez bir do\u011frulama setine girdi. UKVH’ler i\u00e7in, model e\u011fitim s\u00fcrecinde do\u011frulama kesinli\u011fi hakk\u0131nda bilgi sa\u011flamak amac\u0131yla t\u0131rmanma seti olarak hizmet vermek i\u00e7in haz\u0131rl\u0131k verilerinin %10’unu ger\u00e7ekle\u015ftirdi. D\u00fczenlenen setteki her sat\u0131r bir \u00f6\u011frencinin ald\u0131\u011f\u0131 t\u00fcm eylem dizilimlerinden olu\u015fur. Model taraf\u0131ndan \u00fcretilen sonraki do\u011fru eylem tahminlerinin oran\u0131, her \u00f6\u011frenci eylem dizilimi i\u00e7in hesaplan\u0131r. S\u00f6z konusu k\u0131vr\u0131ma6<\/a><\/sup> y\u00f6nelik modelin performans\u0131n\u0131 \u00fcretmek i\u00e7in b\u00fct\u00fcn bir k\u0131vr\u0131m\u0131n oranlar\u0131n\u0131n ortalamas\u0131 al\u0131n\u0131r ve daha sonra belirli bir UKVH model hiperparametre seti i\u00e7in \u00c7D do\u011frulu\u011funu \u00fcretmek \u00fczere be\u015f k\u0131vr\u0131m\u0131n tamam\u0131ndaki performans\u0131n ortalamas\u0131 al\u0131n\u0131r.<\/span><\/p>\n

                                  Y\u00fczeysel Modeller<\/h3>\n

                                  N-gram modeller, basit ama g\u00fc\u00e7l\u00fc olas\u0131l\u0131kl\u0131 modellerdir ve dizilimlerin yap\u0131s\u0131n\u0131 gram olarak adland\u0131r\u0131lan n-boyutlu alt dizilerin istatistikleri yoluyla yakalamay\u0131 hedefler ve n-s\u0131ral\u0131 Markov zincirlerine e\u015fde\u011ferdir. Spesifik olarak, model, xi’nin e\u011fitim setinde \u00f6nceki n-1 durumlar\u0131n\u0131 takip etme olas\u0131l\u0131\u011f\u0131 olan tahmini ko\u015fullu olas\u0131l\u0131k P(xi<\/sub>|i<\/sub>xi<\/sub>_(n_1)<\/sub>, xi\u20131<\/sub>), kullanarak her bir dizi durumunu tahmin eder. N-gram modeller hem h\u0131zl\u0131 ve basit hesaplan\u0131r ve do\u011frudan yorumlara sahiptir. Eylem alan\u0131ndaki olas\u0131 her eylem i\u00e7in bir parametre atayan nispeten y\u00fcksek parametre modelleri olduklar\u0131ndan, n gamlar\u0131n olduk\u00e7a rekabet\u00e7i bir standart olmas\u0131n\u0131 bekliyoruz.<\/span><\/p>\n

                                  N-gram modellerinde, n’nin 2 ila 10 aras\u0131nda de\u011fi\u015fti\u011fini de\u011ferlendirdik; bunlar\u0131n en b\u00fcy\u00fc\u011f\u00fc, e\u011fitim s\u0131ras\u0131nda UKVH ba\u011flam penceresinin boyutuna kar\u015f\u0131l\u0131k gelmektedir. E\u011fitim setinin hi\u00e7bir g\u00f6zlem i\u00e7ermeyen tahminlerini ele almak i\u00e7in, en az bir g\u00f6zlem i\u00e7eren en b\u00fcy\u00fck n-gram\u0131n tahminine tekrar tekrar d\u00f6nmeye dayanan bir y\u00f6ntem olan gerilemeyi7<\/a><\/sup> kulland\u0131k. Do\u011frulama stratejimiz, UKVH modelleriyle ayn\u0131yd\u0131, burada ayn\u0131 be\u015f kat\u0131n ortalama \u00e7apraz do\u011frulama puan\u0131 her model i\u00e7in hesapland\u0131.<\/span><\/p>\n

                                  Ders Yap\u0131s\u0131 Modelleri<\/h3>\n

                                  Dizlim verilerinin varsay\u0131msal yap\u0131sal \u00f6zelliklerinden yararlanmaya y\u00f6nelik \u00e7e\u015fitli alternatif modelleri de d\u00e2hil ettik. Dizilimleri incelerken fark etti\u011fimiz ilk \u015fey, belirli eylemlerin arka arkaya birka\u00e7 kez tekrarlanmas\u0131yd\u0131. Bu nedenle, bu varsay\u0131m\u0131n tek ba\u015f\u0131na veri k\u00fcmesindeki bir sonraki eylemi ne kadar iyi tahmin edebilece\u011fini bilmek \u00f6nemlidir. Daha sonra, ders i\u00e7eri\u011fi en s\u0131k sabit bir dizilimde d\u00fczenlendi\u011finden, ders izlencesinin bir sonraki sayfay\u0131 veya eylemi tahmin etme yetene\u011fini de\u011ferlendirdik. Bunu ders i\u00e7eri\u011findeki sayfalar\u0131, eylem setimizdeki \u00f6\u011frenci sayfa ge\u00e7i\u015fleri ile e\u015fle\u015ftirerek, ders izlencesindeki toplam 300 maddeden 174’\u00fcn\u00fcn e\u015fle\u015fmesi ile sonu\u00e7land\u0131rarak ba\u015fard\u0131k. Eylem alan\u0131m\u0131zda her zaman bulunmayan i\u00e7erik kimli\u011fi dizgilerini e\u015fle\u015ftirmeye g\u00fcvendi\u011fimizden, k\u00fc\u00e7\u00fck bir \u00fcst \u00fcste binen eylemler alt k\u00fcmesi e\u015fle\u015ftirilmedi. Son olarak, mevcut durumun ders program\u0131 i\u00e7inde olmamas\u0131 durumunda, mevcut durumun bir sonraki durum olarak \u00f6ng\u00f6r\u00fclmesi y\u00f6n\u00fcnden her iki modeli de birle\u015ftirdik.<\/span><\/p>\n

                                  SONU\u00c7LAR<\/h2>\n

                                  Bu b\u00f6l\u00fcmde, farkl\u0131 \u00f6\u011frenme oranlar\u0131, katman ba\u015f\u0131na gizli d\u00fc\u011f\u00fcm say\u0131s\u0131 ve UKVH katman say\u0131s\u0131 ile e\u011fitilmi\u015f, daha \u00f6nce bahsedilen UKVH modellerinin sonu\u00e7lar\u0131n\u0131 tart\u0131\u015f\u0131yoruz. Model ba\u015far\u0131s\u0131, 5 kat \u00e7apraz do\u011frulama ile belirlenir ve modelin bir sonraki eylemi ne kadar iyi tahmin etti\u011fi ile ilintilidir. N-gram modelleri ve di\u011fer rota yap\u0131s\u0131 modelleri, 5 kat \u00e7apraz do\u011frulama ile do\u011frulan\u0131r.<\/span><\/p>\n

                                  UKVH Modelleri<\/h3>\n

                                  Tablo 19.3, 10 yeti\u015ftirme s\u00fcresinden sonra hesaplanan t\u00fcm 24 UKVH modelinin \u00c7D do\u011frulu\u011funu g\u00f6stermektedir. .01 \u00f6\u011frenme h\u0131z\u0131na sahip modeller i\u00e7in, tepe t\u0131rmanma setlerinde do\u011fruluk genellikle yineleme 10’da zirveye ula\u015ft\u0131. D\u00fc\u015f\u00fck \u00f6\u011frenme oranlar\u0131na sahip modeller i\u00e7in, en y\u00fcksek \u00c7D do\u011fruluklar\u0131n\u0131n daha fazla e\u011fitimle geli\u015fmesini beklemek makul olacakt\u0131r. Bu modellerin e\u011fitim s\u00fcrecinde ne kadar iyi performans g\u00f6sterdi\u011fine dair bir anl\u0131k g\u00f6r\u00fcnt\u00fc sa\u011flamak yerine 10 tekrardan sonra sonu\u00e7lar\u0131 basit\u00e7e rapor etmeyi se\u00e7tik. Ayr\u0131ca, uzun vadede model performans\u0131n\u0131n.01 \u00f6\u011frenme oran\u0131 model performanslar\u0131 \u00fczerinde ciddi oranda bir geli\u015fme g\u00f6sterme ihtimalinin olmad\u0131\u011f\u0131n\u0131 ve s\u0131n\u0131rl\u0131 GPU hesaplama kaynaklar\u0131nda \u00e7al\u0131\u015facak en umut verici ke\u015fifleri en \u00fcst d\u00fczeye \u00e7\u0131karmam\u0131z gerekti\u011fini varsay\u0131yoruz. Her \u00f6\u011frenme oran\u0131 i\u00e7in en iyi \u00c7D do\u011frulu\u011fu vurgulamak i\u00e7in koyu renkli yap\u0131lm\u0131\u015ft\u0131r.<\/span><\/p>\n

                                  UKVH’leri kullanman\u0131n bir dezavantaj\u0131, bir GPU’ya ihtiya\u00e7 duymalar\u0131 ve e\u011fitilmelerinin nispeten yava\u015f olmas\u0131d\u0131r. Bu nedenle, kullan\u0131lacak en iyi hiperparametreleri ara\u015ft\u0131r\u0131rken, yaln\u0131zca ilk ke\u015fiflerin bir alt k\u00fcmesini temel alan ek modeller e\u011fitmeyi se\u00e7tik. Ayr\u0131ca, ge\u00e7mi\u015f ba\u011flam\u0131 10 \u00f6geden 100 \u00f6geye geni\u015fleterek modele maruz kalan ba\u011flam\u0131 artt\u0131rd\u0131k. Tablo 4, bu geni\u015fletilmi\u015f sonu\u00e7lar\u0131 g\u00f6stermektedir. Her UKVH katman\u0131 256 d\u00fc\u011f\u00fcme sahiptir ve \u00f6nceki hiperparametre arama sonu\u00e7lar\u0131ndaki 10 evre yerine, 20 veya 60 evre i\u00e7in e\u011fitilmi\u015ftir. Geni\u015fletilmi\u015f sonu\u00e7lar, \u00f6nceki sonu\u00e7lara g\u00f6re b\u00fcy\u00fck bir iyile\u015fme g\u00f6stermekte olup, yeni do\u011fruluk, .7093’e k\u0131yasla .7223’te zirveye ula\u015fm\u0131\u015ft\u0131r.<\/span><\/p>\n

                                  \u015eekil 19.3, ilk ke\u015fiften itibaren 1 ve 2 katmanl\u0131 modeller i\u00e7in evreli e\u011fitim s\u0131ras\u0131ndaki %10 tepe t\u0131rmanma hold out k\u00fcmesinin do\u011frulama kesinli\u011fini g\u00f6stermektedir. Her veri noktas\u0131, belirli bir katman ve d\u00fc\u011f\u00fcm say\u0131s\u0131 kombinasyonu i\u00e7in her \u00fc\u00e7 \u00f6\u011frenme h\u0131z\u0131ndaki ortalama t\u0131rmanma do\u011frulu\u011funu temsil eder. Ampirik olarak, daha fazla say\u0131da d\u00fc\u011f\u00fcme sahip olmak, ilk 10 evrede daha y\u00fcksek bir do\u011frulukla ili\u015fkilendirilirken, 2 katmanl\u0131 modeller, kar\u015f\u0131l\u0131k gelen 1 katmanl\u0131 modele yakla\u015fmadan veya ondan \u00f6nce, birka\u00e7 evre i\u00e7in d\u00fc\u015f\u00fck do\u011frulama kesinlikleriyle ba\u015flar. Bu rakam ilk 10 evre i\u00e7in bir anl\u0131k g\u00f6r\u00fcnt\u00fc sa\u011flar; a\u00e7\u0131k\u00e7as\u0131 baz\u0131 parametre kombinasyonlar\u0131 i\u00e7in, daha fazla evre, ek geni\u015fletilmi\u015f UKVH aramas\u0131yla g\u00f6sterildi\u011fi gibi daha y\u00fcksek bir tepe t\u0131rmanma do\u011frulu\u011funa neden olacakt\u0131r. Tahmini olarak, 3 katmanl\u0131 modeller de 2 katmanl\u0131 modellerin sergiledi\u011fi, do\u011fruluklar\u0131n daha alt katman emsallerine k\u0131yasla ba\u015flang\u0131\u00e7ta daha d\u00fc\u015f\u00fck ba\u015flayabildi\u011fi bir e\u011filimi izleyebilir.<\/span><\/p>\n

                                  <\/a><\/a><\/a><\/a><\/a> Tablo 19.3. UKVH Performans\u0131 (10 Evre)<\/i><\/span><\/span><\/p>\n\n\n\n\n\n<\/colgroup>\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
                                  \n

                                  \u00d6\u011frenme Oran\u0131<\/span><\/p>\n<\/td>\n

                                  		\n

                                  D\u00fc\u011f\u00fcmler<\/span><\/p>\n<\/td>\n

                                  		\n

                                  Katmanlar<\/span><\/p>\n<\/td>\n

                                  		\n

                                  Do\u011fruluk<\/span><\/p>\n<\/td>\n<\/tr>\n<\/thead>\n

                                  0.01<\/span><\/span><\/td>\n64<\/span><\/span><\/td>\n1<\/span><\/span><\/td>\n0.7014<\/span><\/span><\/td>\n<\/tr>\n
                                  0.01<\/span><\/span><\/td>\n64<\/span><\/span><\/td>\n2<\/span><\/span><\/td>\n0.7009<\/span><\/span><\/td>\n<\/tr>\n
                                  0.01<\/span><\/span><\/td>\n64<\/span><\/span><\/td>\n3<\/span><\/span><\/td>\n0.6997<\/span><\/span><\/td>\n<\/tr>\n
                                  0.01<\/span><\/span><\/td>\n128<\/span><\/span><\/td>\n1<\/span><\/span><\/td>\n0.7046<\/span><\/span><\/td>\n<\/tr>\n
                                  0.01<\/span><\/span><\/td>\n128<\/span><\/span><\/td>\n2<\/span><\/span><\/td>\n0.7064<\/span><\/span><\/td>\n<\/tr>\n
                                  0.01<\/span><\/span><\/td>\n128<\/span><\/span><\/td>\n3<\/span><\/span><\/td>\n0.7056<\/span><\/span><\/td>\n<\/tr>\n
                                  0.01<\/span><\/span><\/td>\n256<\/span><\/span><\/td>\n1<\/span><\/span><\/td>\n0.7073<\/span><\/span><\/td>\n<\/tr>\n
                                  0.01<\/span><\/span><\/td>\n256<\/span><\/span><\/td>\n2<\/span><\/span><\/td>\n0.7093<\/span><\/span><\/td>\n<\/tr>\n
                                  0.01<\/span><\/span><\/td>\n256<\/span><\/span><\/td>\n3<\/span><\/span><\/td>\n0.7092<\/span><\/span><\/td>\n<\/tr>\n
                                  0.001<\/span><\/span><\/td>\n64<\/span><\/span><\/td>\n1<\/span><\/span><\/td>\n0.6941<\/span><\/span><\/td>\n<\/tr>\n
                                  0.001<\/span><\/span><\/td>\n64<\/span><\/span><\/td>\n2<\/span><\/span><\/td>\n0.6968<\/span><\/span><\/td>\n<\/tr>\n
                                  0.001<\/span><\/span><\/td>\n64<\/span><\/span><\/td>\n3<\/span><\/span><\/td>\n0.6971<\/span><\/span><\/td>\n<\/tr>\n
                                  0.001<\/span><\/span><\/td>\n128<\/span><\/span><\/td>\n1<\/span><\/span><\/td>\n0.6994<\/span><\/span><\/td>\n<\/tr>\n
                                  0.001<\/span><\/span><\/td>\n128<\/span><\/span><\/td>\n2<\/span><\/span><\/td>\n0.7022<\/span><\/span><\/td>\n<\/tr>\n
                                  0.001<\/span><\/span><\/td>\n128<\/span><\/span><\/td>\n3<\/span><\/span><\/td>\n0.7026<\/span><\/span><\/td>\n<\/tr>\n
                                  0.001<\/span><\/span><\/td>\n256<\/span><\/span><\/td>\n1<\/span><\/span><\/td>\n0.7004<\/span><\/span><\/td>\n<\/tr>\n
                                  0.001<\/span><\/span><\/td>\n256<\/span><\/span><\/td>\n2<\/span><\/span><\/td>\n0.7050<\/span><\/span><\/td>\n<\/tr>\n
                                  0.001<\/span><\/span><\/td>\n256<\/span><\/span><\/td>\n3<\/span><\/span><\/td>\n0.7050<\/span><\/span><\/td>\n<\/tr>\n
                                  0.0001<\/span><\/span><\/td>\n64<\/span><\/span><\/td>\n1<\/span><\/span><\/td>\n0.6401<\/span><\/span><\/td>\n<\/tr>\n
                                  0.0001<\/span><\/span><\/td>\n64<\/span><\/span><\/td>\n2<\/span><\/span><\/td>\n0.4719<\/span><\/span><\/td>\n<\/tr>\n
                                  0.0001<\/span><\/span><\/td>\n128<\/span><\/span><\/td>\n1<\/span><\/span><\/td>\n0.6539<\/span><\/span><\/td>\n<\/tr>\n
                                  0.0001<\/span><\/span><\/td>\n128<\/span><\/span><\/td>\n2<\/span><\/span><\/td>\n0.6648<\/span><\/span><\/td>\n<\/tr>\n
                                  0.0001<\/span><\/span><\/td>\n256<\/span><\/span><\/td>\n1<\/span><\/span><\/td>\n0.6677<\/span><\/span><\/td>\n<\/tr>\n
                                  0.0001<\/span><\/span><\/td>\n256<\/span><\/span><\/td>\n2<\/span><\/span><\/td>\n0.6894<\/span><\/span><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

                                   <\/p>\n

                                  <\/a><\/a><\/a><\/a><\/a> Tablo 19.4. Geni\u015fletilmi\u015f UKVH Performans\u0131 (256 D\u00fc\u011f\u00fcm, 100 Pencere Boyutu) <\/i><\/span><\/span><\/p>\n\n\n\n\n\n<\/colgroup>\n\n\n\n\n\n\n\n\n\n\n
                                  \n

                                  \u00d6\u011frenme Oran\u0131<\/span><\/p>\n<\/td>\n

                                  		\n

                                  Evre<\/span><\/p>\n<\/td>\n

                                  		\n

                                  Katmanlar<\/span><\/p>\n<\/td>\n

                                  		\n

                                  Do\u011fruluk<\/span><\/p>\n<\/td>\n<\/tr>\n

                                  0.01<\/span><\/span><\/td>\n20<\/span><\/span><\/td>\n2<\/span><\/span><\/td>\n0.7190<\/span><\/span><\/td>\n<\/tr>\n
                                  0.01<\/span><\/span><\/td>\n60<\/span><\/span><\/td>\n2<\/span><\/span><\/td>\n0.7220<\/span><\/span><\/td>\n<\/tr>\n
                                  0.01<\/span><\/span><\/td>\n20<\/span><\/span><\/td>\n3<\/span><\/span><\/td>\n0.7174<\/span><\/span><\/td>\n<\/tr>\n
                                  0.01<\/span><\/span><\/td>\n60<\/span><\/span><\/td>\n3<\/span><\/span><\/td>\n0.7223<\/span><\/span><\/td>\n<\/tr>\n
                                  0.001<\/span><\/span><\/td>\n20<\/span><\/span><\/td>\n2<\/span><\/span><\/td>\n0.7044<\/span><\/span><\/td>\n<\/tr>\n
                                  0.001<\/span><\/span><\/td>\n60<\/span><\/span><\/td>\n2<\/span><\/span><\/td>\n0.7145<\/span><\/span><\/td>\n<\/tr>\n
                                  0.001<\/span><\/span><\/td>\n20<\/span><\/span><\/td>\n3<\/span><\/span><\/td>\n0.7039<\/span><\/span><\/td>\n<\/tr>\n
                                  0.001<\/span><\/span><\/td>\n60<\/span><\/span><\/td>\n3<\/span><\/span><\/td>\n0.7147<\/span><\/span><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

                                  Ders Yap\u0131s\u0131 Modelleri<\/h3>\n

                                  Farkl\u0131 ders yap\u0131 modelleri i\u00e7in model performans\u0131 Tablo 19.5’te g\u00f6sterilmi\u015ftir. Sonu\u00e7lar, UKVH veya n-gram sonu\u00e7lar\u0131 aral\u0131\u011f\u0131nda dura\u011fanl\u0131k (sonuncusuyla ayn\u0131) veya ders i\u00e7eri\u011fi yap\u0131s\u0131 gibi basit sezgisel taramalardan ya da her iki bulu\u015fsal y\u00f6ntemi de i\u00e7eren (“ders program\u0131 + tekrar”) bir\u00e7ok eylemin tahmin edilebilece\u011fini d\u00fc\u015f\u00fcnd\u00fcrmektedir.<\/span><\/p>\n

                                  \"\"<\/p>\n

                                  <\/a> \u015eekil 19.3. Her bir e\u011fitim k\u00fcmesinin %10’unu olu\u015fturan tepe t\u0131rmanma verilerinde evreye g\u00f6re ortalama do\u011fruluk.<\/i><\/span><\/span><\/p>\n

                                  <\/a><\/a><\/a><\/a><\/a> Tablo 19.5. Yap\u0131sal modeller<\/i><\/span><\/span><\/p>\n\n\n\n<\/colgroup>\n\n\n\n\n\n
                                  \n

                                  Yap\u0131sal model<\/b><\/span><\/p>\n<\/td>\n

                                  		\n

                                  Do\u011fruluk<\/b><\/span><\/p>\n<\/td>\n<\/tr>\n

                                  tekrarlama<\/span><\/span><\/td>\n0.2908<\/span><\/span><\/td>\n<\/tr>\n
                                  \u00f6\u011fretim izlencesi<\/span><\/span><\/td>\n0.2339<\/span><\/span><\/td>\n<\/tr>\n
                                  \u00f6\u011fretim izlencesi + tekrar<\/span><\/span><\/td>\n0.4533<\/span><\/span><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n
                                    \n
                                  1. \n
                                      \n
                                    1. \n
                                        \n
                                      1. \n
                                          \n
                                        1. \n

                                          N-gram Modelleri<\/h4>\n<\/li>\n<\/ol>\n<\/li>\n<\/ol>\n<\/li>\n<\/ol>\n<\/li>\n<\/ol>\n

                                          Model performans\u0131 Tablo 19.6’da g\u00f6sterilmi\u015ftir. En iyi performans g\u00f6steren modeller, \u00f6nceki 7 veya 8 eylemi kullanarak (s\u0131ras\u0131yla 8 ve 9 gram) tahminlerde bulundu. Performans\u0131 art\u0131rmayan daha geni\u015f kay\u0131tlar n aral\u0131\u011f\u0131m\u0131z\u0131n yeterince b\u00fcy\u00fck oldu\u011funu g\u00f6stermi\u015ftir. Genel olarak performans, en iyi n-gram modelin en iyi UKVH modellerinden daha k\u00f6t\u00fc \u00e7al\u0131\u015fmas\u0131na ra\u011fmen n-gram modellerinin UKVH modelleriyle rekabet etti\u011fini g\u00f6stermektedir. Tablo 19.7, en karma\u015f\u0131k model (10 gram) i\u00e7in kullan\u0131lan n-gram modellerin oran\u0131n\u0131 g\u00f6stermektedir. Tahminlerin %62’sinden fazlas\u0131, 10 graml\u0131k g\u00f6zlemler kullan\u0131larak yap\u0131lm\u0131\u015ft\u0131r. Ayr\u0131ca, vakalar\u0131n %1’inden az\u0131 tahminleri yapmak i\u00e7in unigramlara veya bigramlara geri d\u00f6nd\u00fc ve bu da daha b\u00fcy\u00fck gram \u00f6r\u00fcnt\u00fcleri i\u00e7in \u00f6nemli bir g\u00f6zlem eksikli\u011fi olmad\u0131\u011f\u0131n\u0131 \u00f6ne s\u00fcrd\u00fc.<\/span><\/p>\n

                                          <\/a><\/a><\/a><\/a><\/a> Tablo 19.6. N-gram Performans<\/i><\/span><\/span><\/p>\n\n\n\n<\/colgroup>\n\n\n\n\n\n\n\n\n\n\n\n
                                          \n

                                          N-gram<\/span><\/p>\n<\/td>\n

                                          		\n

                                          Do\u011fruluk<\/span><\/p>\n<\/td>\n<\/tr>\n

                                          2-gram<\/span><\/span><\/td>\n0.6304<\/span><\/span><\/td>\n<\/tr>\n
                                          3 gram<\/span><\/span><\/td>\n0.6658<\/span><\/span><\/td>\n<\/tr>\n
                                          4 gram<\/span><\/span><\/td>\n0.6893<\/span><\/span><\/td>\n<\/tr>\n
                                          5 gram<\/span><\/span><\/td>\n0.6969<\/span><\/span><\/td>\n<\/tr>\n
                                          6 gram<\/span><\/span><\/td>\n0.7012<\/span><\/span><\/td>\n<\/tr>\n
                                          7 gram<\/span><\/span><\/td>\n0.7030<\/span><\/span><\/td>\n<\/tr>\n
                                          8 gram<\/span><\/span><\/td>\n0.7035<\/span><\/span><\/td>\n<\/tr>\n
                                          9 gram<\/span><\/span><\/td>\n0.7035<\/span><\/span><\/td>\n<\/tr>\n
                                          10 gram<\/span><\/span><\/td>\n0.7033<\/span><\/span><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

                                           <\/p>\n

                                          <\/a><\/a><\/a><\/a><\/a> Tablo 19.7. N cinsinden 10 graml\u0131k \u00f6ng\u00f6r\u00fcn\u00fcn oran\u0131<\/i><\/span><\/span><\/p>\n\n\n\n<\/colgroup>\n\n\n\n\n\n\n\n\n\n\n\n\n\n
                                          \n

                                          n<\/span><\/p>\n<\/td>\n

                                          		\n

                                          \u00d6ng\u00f6r\u00fclen %<\/b><\/span><\/p>\n<\/td>\n<\/tr>\n<\/thead>\n

                                          1 <\/span><\/span><\/td>\n0.0003<\/span><\/span><\/td>\n<\/tr>\n
                                          2<\/span><\/span><\/td>\n0.0084<\/span><\/span><\/td>\n<\/tr>\n
                                          3<\/span><\/span><\/td>\n0.0210<\/span><\/span><\/td>\n<\/tr>\n
                                          4<\/span><\/span><\/td>\n0.0423<\/span><\/span><\/td>\n<\/tr>\n
                                          5<\/span><\/span><\/td>\n0.0524<\/span><\/span><\/td>\n<\/tr>\n
                                          6<\/span><\/span><\/td>\n0.0605<\/span><\/span><\/td>\n<\/tr>\n
                                          7<\/span><\/span><\/td>\n0.0624<\/span><\/span><\/td>\n<\/tr>\n
                                          8<\/span><\/span><\/td>\n0.0615<\/span><\/span><\/td>\n<\/tr>\n
                                          9<\/span><\/span><\/td>\n0.0594<\/span><\/span><\/td>\n<\/tr>\n
                                          10<\/span><\/span><\/td>\n0.6229<\/span><\/span><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

                                          Yine de yakla\u015f\u0131k %6 daha az veri noktas\u0131 art arda gelen daha b\u00fcy\u00fck n-gramlar taraf\u0131ndan tahmin ediliyor gibi g\u00f6r\u00fcnmektedir.<\/span><\/p>\n

                                            \n
                                          1. \n
                                              \n
                                            1. \n
                                                \n
                                              1. \n
                                                  \n
                                                1. \n

                                                  Sertifikas\u0131z \u00d6\u011frencilerin Do\u011frulanmas\u0131<\/h4>\n<\/li>\n<\/ol>\n<\/li>\n<\/ol>\n<\/li>\n<\/ol>\n<\/li>\n<\/ol>\n

                                                  Sonunda sertifika almayan \u00f6\u011frencilerden gelen veri ak\u0131\u015flar\u0131 \u00fczerindeki eylemleri tahmin etmek i\u00e7in 10 evreli e\u011fitimden sonra (.01 \u00f6\u011frenme h\u0131z\u0131, 256 d\u00fc\u011f\u00fcm, iki katman) en iyi performans g\u00f6steren \u201corijinal\u201d UKVH modelini kulland\u0131k. Sertifikas\u0131z \u00f6\u011frencilerin \u00e7o\u011fu yaln\u0131zca birka\u00e7 oturum a\u00e7ma eylemi ger\u00e7ekle\u015ftirdi, bu nedenle analizi en az 30 oturum a\u00e7ma eylemi olan \u00f6\u011frencilerle s\u0131n\u0131rlad\u0131k. Bu kriterleri kar\u015f\u0131layan 10.761 \u00f6\u011frenci ve toplam 2.151.666 eylem vard\u0131. UKVH modeli, sertifikal\u0131 \u00f6\u011frenciler i\u00e7in .7093 \u00e7apraz do\u011frulanm\u0131\u015f do\u011frulukla kar\u015f\u0131la\u015ft\u0131r\u0131ld\u0131\u011f\u0131nda, do\u011frulanmam\u0131\u015f \u00f6\u011frenci alan\u0131ndan gelen eylemleri .6709 do\u011frulukla tam bir \u015fekilde tahmin edebildi. Bu fark, sertifikal\u0131 \u00f6\u011frencilerden gelen eylemlerin, belgelendirilmemi\u015f \u00f6\u011frencilerden gelen eylemlerden farkl\u0131 olma e\u011filiminde oldu\u011funu g\u00f6sterir, belki de \u00f6\u011frencilere rehberlik etmek i\u00e7in otomatik bir \u00f6neri \u00e7er\u00e7evesi sa\u011flamada potansiyel bir uygulama g\u00f6stermektedir.<\/span><\/p>\n

                                                  <\/a> Tablo 19.8. \u00c7apraz Do\u011frulanm\u0131\u015f Modellerin Kar\u015f\u0131la\u015ft\u0131r\u0131lmas\u0131<\/i><\/span><\/span><\/p>\n\n\n\n\n<\/colgroup>\n\n\n\n\n
                                                  \n

                                                  N-gram<\/b><\/span><\/p>\n<\/td>\n

                                                  		\n

                                                  Do\u011fru<\/b><\/span><\/p>\n<\/td>\n

                                                  		\n

                                                  N-gram Yanl\u0131\u015f<\/b><\/span><\/p>\n<\/td>\n<\/tr>\n

                                                  UKVH Do\u011fru<\/span><\/span><\/td>\n7.565.862<\/span><\/span><\/td>\n577.683<\/span><\/span><\/td>\n<\/tr>\n
                                                  UKVH Yanl\u0131\u015f<\/span><\/span><\/td>\n367.960<\/span><\/span><\/td>\n2.735.702<\/span><\/span><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

                                                  KATKILAR<\/h2>\n

                                                  Bu \u00e7al\u0131\u015fmada, gran\u00fcler \u00f6\u011frenci eylem verilerinin modellenmesi sorununa bir KA\u00c7D i\u00e7indeki her t\u00fcr etkile\u015fimi modelleyerek yakla\u015ft\u0131k. Bu \u00f6ncelikli olarak de\u011ferlendirme sonu\u00e7lar\u0131n\u0131 kullanarak gizli \u00f6\u011frenci bilgisini modellemeye odaklanan \u00f6nceki \u00e7al\u0131\u015fmalardan farkl\u0131d\u0131r. Bir \u00f6\u011frencinin bir sonraki eylemini tahmin ederken, en iyi performans g\u00f6steren UKVH modeli .7223 \u00e7apraz do\u011frulama kesinli\u011fi \u00fcretti ki bu da en iyi n-gram model do\u011frulu\u011fu olan .7035’in \u00fczerinde bir geli\u015fme olarak toplam 11 milyon olas\u0131 tahminden 210.000 daha do\u011fru tahmindir. Tablo 19.8, iki modelin \u00e7apraz onaylama s\u0131ras\u0131nda do\u011fru veya yanl\u0131\u015f bir \u00f6ng\u00f6r\u00fcde mutab\u0131k kald\u0131\u011f\u0131 veya kalmad\u0131\u011f\u0131 say\u0131y\u0131 g\u00f6stermektedir. Hem UKVH hem de n-gram modelleri, bir sonraki eylemin \u00f6\u011fretim program\u0131 izlencesi yap\u0131s\u0131 ve tekrarlar arac\u0131l\u0131\u011f\u0131yla \u00f6ng\u00f6r\u00fclmesi yap\u0131sal modeli \u00fczerinde \u00f6nemli bir geli\u015fme sa\u011flar; bu, \u00f6\u011frenci kat\u0131l\u0131m \u00f6r\u00fcnt\u00fclerinin ders materyali i\u00e7erisinde tamamen do\u011frusal bir gezinmeden a\u00e7\u0131k\u00e7a sapt\u0131\u011f\u0131n\u0131 g\u00f6sterir.<\/span><\/p>\n

                                                  Bildi\u011fimiz kadar\u0131yla, bu b\u00f6l\u00fcm bir KA\u00c7D’de davran\u0131\u015fsal verinin bu ayr\u0131nt\u0131 d\u00fczeyinde tahmin edildi\u011fini ilk kez g\u00f6stermektedir. Ayr\u0131ca, KA\u00c7D verilerine ilk defa tekrarlayan sinir a\u011flar\u0131n\u0131n uyguland\u0131\u011f\u0131n\u0131 g\u00f6sterir. Bu tekni\u011fin \u00f6\u011frenci davran\u0131\u015fsal durumlar\u0131n\u0131 ham zaman serisi verileri ile temsil etmek i\u00e7in, \u00f6zellik m\u00fchendisli\u011fi olmadan, y\u00fcksek hacimli zaman serisi verileri ile herhangi bir \u00f6\u011frenme analiti\u011fi ba\u011flam\u0131nda geni\u015f bir uygulanabilirli\u011fe sahip oldu\u011funa inan\u0131yoruz. \u00c7er\u00e7evelememiz, davran\u0131\u015fsal veri modellerinin \u00f6\u011frenci i\u00e7in gelecekteki davran\u0131\u015flar\u0131 \u00f6nermek i\u00e7in nas\u0131l kullan\u0131labilece\u011fini ortaya koyarken, davran\u0131\u015fsal durumlar\u0131n\u0131n temsili, performanstan duyu\u015fsal duruma kadar \u00e7e\u015fitli yap\u0131larda \u00e7e\u015fitli \u00e7\u0131kar\u0131mlar yapmak i\u00e7in de\u011ferli olabilir.<\/span><\/p>\n

                                                  GELECEKTE YAPILACAKLAR<\/h2>\n

                                                  Hem UKVH hem de n-gram modelleri de geli\u015ftirilebilir. \u00d6zellikle, n-gram modellerimiz g\u00f6r\u00fcnmeyen gramlar\u0131n daha iyi kullan\u0131lmas\u0131na izin veren geri tepme ve p\u00fcr\u00fczs\u00fczle\u015ftirme tekniklerinin bir kombinasyonundan faydalanabilir. UKVH, daha geni\u015f bir hiperparametre g\u00f6zene\u011fi arama, daha fazla yeti\u015ftirme s\u00fcresi, daha uzun yeti\u015ftirme ba\u011flam\u0131 pencereleri ve daha y\u00fcksek boyutlu eylem yerle\u015ftirmelerinden faydalanabilir. Ek olarak, veri k\u00fcmemizdeki sinyal-g\u00fcr\u00fclt\u00fc oran\u0131, daha az bilgilendirici veya gereksiz \u00f6\u011frenci eylemleri kald\u0131r\u0131larak veya eylemler aras\u0131ndaki s\u00fcreyi temsil etmek i\u00e7in ek belirte\u00e7ler eklenerek artt\u0131r\u0131labilir.<\/span><\/p>\n

                                                  Derin \u00f6\u011frenme modellerini b\u00fcy\u00fck \u00f6\u011frenci eylemi veri k\u00fcmelerine uygulaman\u0131n birincil nedeni, KA\u00c7D ortamlar\u0131ndaki \u00f6\u011frenci davran\u0131\u015f\u0131n\u0131 modellemektir. Bu \u00f6r\u00fcnt\u00fcler, otomatikle\u015ftirilmi\u015f \u00f6neri sistemlerinin olu\u015fturulmas\u0131na yard\u0131mc\u0131 olmak i\u00e7in kullan\u0131labilir; burada, zorlu bir \u00f6\u011frenciye, ge\u00e7mi\u015f davran\u0131\u015flar\u0131na ve performanslar\u0131na g\u00f6re i\u00e7eri\u011fi g\u00f6r\u00fcnt\u00fclemek i\u00e7in ge\u00e7i\u015f \u00f6nerileri sa\u011flanabilir. B\u00f6yle bir uygulaman\u0131n olas\u0131l\u0131\u011f\u0131n\u0131 de\u011ferlendirmek i\u00e7in, a\u011f\u0131m\u0131zdan t\u00fcretilmi\u015f bir \u00f6neri sistemini, y\u00f6nlendirilmemi\u015f bir kontrol grubuna kar\u015f\u0131 deneysel olarak test etmeyi planl\u0131yoruz. Ek olarak, gelecekteki \u00e7al\u0131\u015fmalar, \u00e7e\u015fitli dersler i\u00e7in benzer modellerin performans\u0131n\u0131 de\u011ferlendirmeli ve tek bir model kullanarak genel ders \u00f6r\u00fcnt\u00fclerinin ne \u00f6l\u00e7\u00fcde \u00f6\u011frenilebilece\u011fini incelemelidir. Bu b\u00f6l\u00fcmde \u00f6nerilen modeller, bilgi i\u015flemsel bir davran\u0131\u015f modelini s\u00fcrd\u00fcrmektedir. KA\u00c7D’lerdeki \u00f6\u011frenci davran\u0131\u015f dizilimlerinde d\u00fczenliliklerin var oldu\u011fu bu modeller arac\u0131l\u0131\u011f\u0131yla g\u00f6sterilmi\u015ftir. Bilgi i\u015flemsel bir modelin bu kal\u0131plar\u0131 saptayabildi\u011fi g\u00f6z \u00f6n\u00fcne al\u0131nd\u0131\u011f\u0131nda, model bize \u00f6\u011frenci davran\u0131\u015flar\u0131 hakk\u0131nda daha geni\u015f kapsaml\u0131 ne s\u00f6yleyebilir ve bu bulgular mevcut davran\u0131\u015f teorileriyle nas\u0131l ba\u011flant\u0131 kurabilir ve bunlar\u0131 nas\u0131l kurabilir? Her zaman diliminde model \u00f6\u011frenci i\u00e7in gizli bir davran\u0131\u015f durumunu takip etti\u011finden, bu durum o anda mevcut oldu\u011fu bilinen \u00f6\u011frencilerin di\u011fer \u00f6zellikleri ile g\u00f6rselle\u015ftirilebilir ve ili\u015fkilendirilebilir. Gelecekteki \u00e7al\u0131\u015fmalar, \u00f6\u011frencinin durumu hakk\u0131ndaki kendi anlay\u0131\u015f\u0131m\u0131z\u0131 bilgilendirmeye yard\u0131mc\u0131 olabilmesi i\u00e7in bu bilgi i\u015flemsel davran\u0131\u015f modelini geli\u015ftirmeye \u00e7al\u0131\u015facakt\u0131r.<\/span><\/p>\n

                                                  TE\u015eEKK\u00dcR B\u00d6L\u00dcM\u00dc<\/h2>\n

                                                  Bu \u00e7al\u0131\u015fma, Ulusal Bilim Vakf\u0131’ndan bir hibe ile desteklenmi\u015ftir (IIS: BIGDATA 1547055).<\/span><\/p>\n

                                                  KAYNAK\u00c7A<\/h2>\n

                                                  Bastien, F., Lamblin, P., Pascanu, R., Bergstra, J., Goodfellow, I. J., Bergeron, A., . . . Bengio, Y. (2012). Theano: New features and speed improvements. Deep Learning and Unsupervised Feature Learning NIPS 2012 Workshop. Advances in Neural Information Processing Systems 25 <\/i>(NIPS 2012), 3\u20138 December 2012, Lake Tahoe, NV, USA. http:\/\/www.iro.umontreal.ca\/~lisa\/pointeurs\/nips2012_deep_workshop_theano_final.pdf <\/span><\/span><\/p>\n

                                                  Bengio, Y., Simard, P., & Frasconi, P. (1994). Learning long-term dependencies with gradient descent is difficult. IEEE Transactions on Neural Networks, 5<\/i>(2), 157\u2013166. <\/span><\/span><\/p>\n

                                                  Bergstra, J., Breuleux, O., Bastien, F., Lamblin, P., Pascanu, R., Desjardins, G., . . . Bengio, Y. (2010, June). Theano: A CPU and GPU math expression compiler. Proceedings of the Python for Scientific Computing Conference <\/i>(SciPy 2010), 28 June\u20133 July 2010, Austin, TX, USA (pp. 3\u201310).<\/span><\/span><\/p>\n

                                                  Brown, P. F., Desouza, P. V., Mercer, R. L., Pietra, V. J. D., & Lai, J. C. (1992). Class-based n-gram models of natural language. Computational Linguistics, 18<\/i>(4), 467\u2013479. <\/span><\/span><\/p>\n

                                                  Chollet, F. (2015). Keras. GitHub. https:\/\/github.com\/fchollet\/keras <\/span><\/span><\/p>\n

                                                  Corbett, A. T., & Anderson, J. R. (1994). Knowledge tracing: Modeling the acquisition of procedural knowledge. User Modeling and User-Adapted Interaction, 4<\/i>(4), 253\u2013278. <\/span><\/span><\/p>\n

                                                  Crossley, S., Paquette, L., Dascalu, M., McNamara, D. S., & Baker, R. S. (2016). Combining click-stream data with NLP tools to better understand MOOC completion. Proceedings of the 6th International Conference on Learning Analytics and Knowledge <\/i>(LAK\u201916), 25\u201329 April 2016, Edinburgh, UK (pp. 6\u201314). New York: ACM. <\/span><\/span><\/p>\n

                                                  Gers, F. A., Schmidhuber, J., & Cummins, F. (2000). Learning to forget: Continual prediction with LSTM. Neural Computation, 12<\/i>(10), 2451\u20132471. <\/span><\/span><\/p>\n

                                                  Goldberg, Y., & Levy, O. (2014). Word2vec explained: Deriving Mikolov et al.\u2019s negative-sampling word-embedding method. CoRR. arxiv.org\/abs\/1402.3722 <\/span><\/span><\/p>\n

                                                  Graves, A., Mohamed, A.-r., & Hinton, G. (2013). Speech recognition with deep recurrent neural networks. Proceedings of the 2013 IEEE International Conference on Acoustics, Speech and Signal Processing <\/i>(ICASSP 2013), 26\u201331 May, Vancouver, BC, Canada (pp. 6645\u20136649). Institute of Electrical and Electronics Engineers. <\/span><\/span><\/p>\n

                                                  Greff, K., Srivastava, R. K., Koutn\u00edk, J., Steunebrink, B. R., & Schmidhuber, J. (2015). LSTM: A search space odyssey. arXiv preprint arXiv:1503.04069. <\/span><\/span><\/p>\n

                                                  Hochreiter, S., & Schmidhuber, J. (1997). Long short-term memory. Neural Computation, 9<\/i>(8), 1735\u20131780. <\/span><\/span><\/p>\n

                                                  Khajah, M., Lindsey, R. V., & Mozer, M. C. (2016). How deep is knowledge tracing? arXiv preprint arXiv:1604.02416. <\/span><\/span><\/p>\n

                                                  Mikolov, T., Karafi\u00e1t, M., Burget, L., Cernocky, J., & Khudanpur, S. (2010). Recurrent neural network based language model. Proceedings of the 11th Annual Conference of the International Speech Communication Association <\/i>(INTERSPEECH 2010), 26\u201330 September 2010, Makuhari, Chiba, Japan (pp. 1045\u20131048). http:\/\/www.fit.vutbr.cz\/research\/groups\/speech\/publi\/2010\/mikolov_interspeech2010_IS100722.pdf <\/span><\/span><\/p>\n

                                                  Oleksandra, P., & Shane, D. (2016). Untangling MOOC learner networks. Proceedings of the 6th International Conference on Learning Analytics and Knowledge <\/i>(LAK\u201916), 25\u201329 April 2016, Edinburgh, UK (pp. 208\u2013212). New York: ACM. <\/span><\/span><\/p>\n

                                                  Pardos, Z. A., Bergner, Y., Seaton, D. T., & Pritchard, D. E. (2013). Adapting Bayesian knowledge tracing to a massive open online course in EDX. In S. K. D\u2019Mello et al. (Eds.), Proceedings of the 6th International Conference on Educational Data Mining <\/i>(EDM2013), 6\u20139 July 2013, Memphis, TN, USA (pp. 137\u2013144). International Educational Data Mining Society\/Springer. <\/span><\/span><\/p>\n

                                                  Pardos, Z. A., & Xu, Y. (2016). Improving efficacy attribution in a self-directed learning environment using prior knowledge individualization. Proceedings of the 6th International Conference on Learning Analytics and Knowledge <\/i>(LAK\u201916), 25\u201329 April 2016, Edinburgh, UK (pp. 435\u2013439). New York: ACM. <\/span><\/span><\/p>\n

                                                  Pham, V., Bluche, T., Kermorvant, C., & Louradour, J. (2014). Dropout improves recurrent neural networks for handwriting recognition. Proceedings of the 14th International Conference on Frontiers in Handwriting Recognition <\/i>(ICFHR 2014) 1\u20134 September 2014, Crete, Greece (pp. 285\u2013290). <\/span><\/span><\/p>\n

                                                  Piech, C., Bassen, J., Huang, J., Ganguli, S., Sahami, M., Guibas, L. J., & Sohl-Dickstein, J. (2015). Deep knowledge tracing. In C. Cortes et al. (Eds.), Advances in Neural Information Processing Systems 28 <\/i>(NIPS 2015), 7\u201312 December 2015, Montreal, QC, Canada (pp. 505\u2013513). <\/span><\/span><\/p>\n

                                                  Reddy, S., Labutov, I., & Joachims, T. (2016). Latent skill embedding for personalized lesson sequence recommendation. CoRR. arxiv.org\/abs\/1602.07029 <\/span><\/span><\/p>\n

                                                  Reich, J., Stewart, B., Mavon, K., & Tingley, D. (2016). The civic mission of MOOCs: Measuring engagement across political differences in forums. Proceedings of the 3rd ACM Conference on Learning @ Scale <\/i>(L@S 2016), 25\u201328 April 2016, Edinburgh, Scotland (pp. 1\u201310). New York: ACM.<\/span><\/span><\/p>\n

                                                  Sharma, A., Biswas, A., Gandhi, A., Patil, S., & Deshmukh, O. (2016). Livelinet: A multimodal deep recurrent neural network to predict liveliness in educational videos. In T. Barnes et al. (Eds.), Proceedings of the 9th International Conference on Educational Data Mining <\/i>(EDM2016), 29 June\u20132 July 2016, Raleigh, NC, USA. International Educational Data Mining Society. http:\/\/www.educationaldatamining.org\/EDM2016\/proceedings\/paper_64.pdf <\/span><\/span><\/p>\n

                                                  Vinyals, O., Kaiser, L. Koo, T., Petrov, S., Sutskever, I., & Hinton, G. (2015). Grammar as a foreign language. In C. Cortes et al. (Eds.), Advances in Neural Information Processing Systems 28 <\/i>(NIPS 2015), 7\u201312 December 2015, Montreal, QC, Canada (pp. 2755\u20132763). http:\/\/papers.nips.cc\/paper\/5635-grammar-as-a-foreign-language.pdf <\/span><\/span><\/p>\n

                                                  Vinyals, O., Toshev, A., Bengio, S., & Erhan, D. (2015). Show and tell: A neural image caption generator. Proceedings of the 2015 IEEE Computer Society Conference on Computer Vision and Pattern Recognition <\/i>(CVPR 2015), 8\u201310 June 2015, Boston, MA, USA. IEEE Computer Society. arXiv:1411.4555 <\/span><\/span><\/p>\n

                                                  Wen, M., & Ros\u00e9, C. P. (2014). Identifying latent study habits by mining learner behavior patterns in massive open online courses. Proceedings of the 23rd ACM International Conference on Information and Knowledge Management <\/i>(CIKM\u201914), 3\u20137 November 2014, Shanghai, China (pp. 1983\u20131986). New York: ACM. <\/span><\/span><\/p>\n

                                                  Wen, M., Yang, D., & Ros\u00e9, C. P. (2014). Sentiment analysis in MOOC discussion forums: What does it tell us? In J. Stamper et al. (Eds.), Proceedings of the 7th International Conference on Educational Data Mining <\/i>(EDM2014), 4\u20137 July 2014, London, UK. International Educational Data Mining Society. http:\/\/www.cs.cmu.edu\/~mwen\/papers\/edm2014-camera-ready.pdf <\/span><\/span><\/p>\n

                                                  Werbos, P. J. (1988). Generalization of backpropagation with application to a recurrent gas market model. Neural Networks, 1<\/i>(4), 339\u2013356. <\/span><\/span><\/p>\n

                                                  Zaremba, W., Sutskever, I., & Vinyals, O. (2014). Recurrent neural network regularization. arXiv:1409.2329.<\/span><\/span><\/p>\n

                                                  \n
                                                  \n

                                                  1<\/a> orj. feature engineering<\/span><\/span><\/p>\n<\/div>\n

                                                  \n

                                                  2<\/a>orj. generative<\/span><\/span><\/p>\n<\/div>\n

                                                  \n

                                                  3<\/a>orj. embedding<\/span><\/span><\/p>\n<\/div>\n

                                                  \n

                                                  4<\/a> https:\/\/github.com\/CAHLR\/mooc-behaviorcase-study<\/span><\/span><\/p>\n<\/div>\n

                                                  \n

                                                  5<\/a> orj. epoch<\/span><\/span><\/p>\n<\/div>\n

                                                  \n

                                                  6<\/a> orj. fold<\/span><\/span><\/p>\n<\/div>\n

                                                  \n

                                                  7<\/a> orj.backoff<\/span><\/span><\/p>\n<\/div>\n","protected":false},"author":1,"menu_order":9,"template":"","meta":{"pb_show_title":"on","pb_short_title":"","pb_subtitle":"","pb_authors":[],"pb_section_license":""},"chapter-type":[48],"contributor":[],"license":[],"class_list":["post-108","chapter","type-chapter","status-publish","hentry","chapter-type-numberless"],"part":73,"_links":{"self":[{"href":"https:\/\/acikkitap.com.tr\/oaek\/wp-json\/pressbooks\/v2\/chapters\/108","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/acikkitap.com.tr\/oaek\/wp-json\/pressbooks\/v2\/chapters"}],"about":[{"href":"https:\/\/acikkitap.com.tr\/oaek\/wp-json\/wp\/v2\/types\/chapter"}],"author":[{"embeddable":true,"href":"https:\/\/acikkitap.com.tr\/oaek\/wp-json\/wp\/v2\/users\/1"}],"version-history":[{"count":0,"href":"https:\/\/acikkitap.com.tr\/oaek\/wp-json\/pressbooks\/v2\/chapters\/108\/revisions"}],"part":[{"href":"https:\/\/acikkitap.com.tr\/oaek\/wp-json\/pressbooks\/v2\/parts\/73"}],"metadata":[{"href":"https:\/\/acikkitap.com.tr\/oaek\/wp-json\/pressbooks\/v2\/chapters\/108\/metadata\/"}],"wp:attachment":[{"href":"https:\/\/acikkitap.com.tr\/oaek\/wp-json\/wp\/v2\/media?parent=108"}],"wp:term":[{"taxonomy":"chapter-type","embeddable":true,"href":"https:\/\/acikkitap.com.tr\/oaek\/wp-json\/pressbooks\/v2\/chapter-type?post=108"},{"taxonomy":"contributor","embeddable":true,"href":"https:\/\/acikkitap.com.tr\/oaek\/wp-json\/wp\/v2\/contributor?post=108"},{"taxonomy":"license","embeddable":true,"href":"https:\/\/acikkitap.com.tr\/oaek\/wp-json\/wp\/v2\/license?post=108"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}