Publication: Derin öğrenme yöntemleri kullanılarak ürün önerme sistemi tasarımı
Abstract
Bu tez çalışmasında, derin öğrenme yöntemi olan Otomatik Kodlayıcılar (AutoEncoder) kullanılarak bir ürün öneri sistemi tasarlanmıştır. Tasarlanan öneri sistemi TensorFlow platformunda Python programlama dili kullanılarak gerçekleştirilmiştir. Kullanılan veri seti, GroupLens araştırmacılarının MovieLens internet sitesi kullanıcılarından topladığı 1 ile 5 arasındaki puanlardan (açık geri bildirim) oluşan MovieLens 1M veri setidir.Yapılan tez çalışması dört bölümden oluşmaktadır. Birinci bölümde öneri sisteminin ve derin öğrenmenin önemi, literatürdeki çalışmalar ve bir öneri sistemi tasarlanırken kullanılan yöntemler açıklanmıştır. İkinci bölümde ise tasarlanan öneri sistemi aşamalar halinde anlatılmıştır. Sistemin başarısının artırılması için en uygun optimizasyon algoritmasının bulunmasına çalışılmış ve veri miktarındaki artışın optimizasyon algoritması üzerindeki etkisi incelenmiştir. İncelenen Gradyan Alçalma, Hızlı Gradyan Alçalma, RmsProp ve Adam (Adaptive Momentum Optimization) algoritmalarının sonuçları tablolar ve grafikler halinde gösterilmiştir. Üçüncü bölümde ise sonuçlar değerlendirilmiş ve Adam algoritmasının 1,363 puan test hatası ile en başarılı algoritma olduğu tespit edilmiştir. Ayrıca eğitim setindeki veri miktarı arttıkça öneri sisteminin başarısının arttığı görülmüştür. Dördüncü bölümde ise bu sonuçlar açıklanmış daha iyi bir öneri sistemi için yapılabilecek ileriye yönelik planlanan çalışmalardan bahsedilmiştir.
In this study, a product recommendation system was designed using the deep learning method AutoEncoder. This designed recommendation system was implemented using the Python language on the TensorFlow platform. MovieLens 1M dataset was used which consists of the movie ratings ranging from 1 to 5 and collected by the GroupLens researchers from the users of MovieLens website. The study was written in four sections. In the first section, the significance of the recommendation system and the deep learning approach, the studies which already exist in the literature, and the methods that are used when designing a recommendation system were clarified while in the second section, the designed recommendation system was elaborated step by step. Furthermore, it was tried to find the best optimization algorithm to improve the success of the system and the effect of increasing the amount of data that was analyzed. The results of the algorithms of Gradient Descent, Gradient Descent with Momentum, RmsProp and Adam were depicted in both tables and graphs. In the third section, findings were evaluated and as a result, the Adam was shown to be the best algorithm with 1,363 points of test error. It was observed that the more data a training set has, the more successful the recommendation system is. In the fourth section, all the results from the processes mentioned above were summarized and the studies which can be conducted in order to implement a better recommendation system as future plans were stated.
In this study, a product recommendation system was designed using the deep learning method AutoEncoder. This designed recommendation system was implemented using the Python language on the TensorFlow platform. MovieLens 1M dataset was used which consists of the movie ratings ranging from 1 to 5 and collected by the GroupLens researchers from the users of MovieLens website. The study was written in four sections. In the first section, the significance of the recommendation system and the deep learning approach, the studies which already exist in the literature, and the methods that are used when designing a recommendation system were clarified while in the second section, the designed recommendation system was elaborated step by step. Furthermore, it was tried to find the best optimization algorithm to improve the success of the system and the effect of increasing the amount of data that was analyzed. The results of the algorithms of Gradient Descent, Gradient Descent with Momentum, RmsProp and Adam were depicted in both tables and graphs. In the third section, findings were evaluated and as a result, the Adam was shown to be the best algorithm with 1,363 points of test error. It was observed that the more data a training set has, the more successful the recommendation system is. In the fourth section, all the results from the processes mentioned above were summarized and the studies which can be conducted in order to implement a better recommendation system as future plans were stated.