Sequential recommenders systems using deep learning for e-commerce sites

Abstract

The recommender system is becoming ubiquitous in the era when an overwhelming amount of information surrounds people. Utilizing effective recommendations has been shown to enhance business and make it simpler for customers to choose from an endless list of options. Especially on e-commerce websites, where many products we buy online are suggested to us by recommender systems. This thesis focuses on an approach known as sequential recommendations, which leverages historical interactions in chronological order to discover user preferences and predict the most referenced products based on the user’s historical behaviour sequence. The applied deep learning models are the Self-attentive Sequential Recommendation (SASRec) and Sequential Recommendation with Bidirectional Encoder Representations from Transformer (BERT4Rec). The performance of the recommender system was evaluated using Amazon Books dataset and measured with ranking and accuracy metrics. Moreover, the model has also been implemented in a web prototype to provide real-time book recommendations to the user and demonstrate practical use cases of the model in the context of an e-commerce site.