EXPLORING DEEP LEARNING TECHNIQUES TO TACKLE THE SPARSITY PROBLEM IN RECOMMENDER SYSTEMS

Abstract

With the inception of e-commerce in the early twenty-first century, people's lifestyles have drastically changed. People today tend to do many of their daily routines online, such as shopping, reading the news, and watching movies. Nevertheless, consumers often face difficulties while exploring related items such as new fashion trends because they are not aware of their existence due to the overwhelming amount of information available online. This phenomenon is widely known as ``information overload''. Therefore, recommender systems (RSs) are a critical solution for helping users make decisions when there are lots of choices. RSs have been integrated into and have become an essential part of every website due to their effectiveness in increasing customer interactions, attracting new customers, and growing business revenue.

Machine learning, and deep learning (DL) in particular, have achieved a great success in resolving various computer science problems. Generally, DL-based approaches have enhanced performance remarkably compared with traditional approaches. Specifically, DL-based approaches have become the state-of-the-art techniques in RSs. Therefore, in this dissertation, three DL-based contributions are presented to address the natural data sparsity problem in RSs: (1) DeepHCF, a deep-hybrid, collaborative-filtering model that trains two deep models via joint training for rating prediction tasks; (2) CATA, a collaborative attentive autoencoder that integrates the attention mechanism to enhance the recommendation quality for ranking prediction tasks; and (3) CATA++, an extended version of CATA that employs a dual attentive autoencoder to leverage more of the item's content. All proposed models have gone through comprehensive experiments to evaluate their performance against state-of-the-art models using real-world datasets. Our experimental results show the superiority of our models over state-of-the-art models according to various evaluation metrics.