Cross-domain recommender systems using deep neural networks

Doctor of PhilosophyDepartment of Computer ScienceDoina CarageaWith the continuous growth in the number of products available in e-commerce applications and information items available on the Internet, the task of associating users with a small list of personalized items, extracted from a large and diverse pool of items, is clearly beyond human ability, a problem known as Information Overload. Indeed, the task (or ability) of automatically, quickly, and accurately recommending appropriate items to users has become an essential part in determining the success of almost all e-commerce businesses and online service providers. Recommender Systems (RS) aim to exploit users' historical interaction records, to capture users' preferences, and accordingly identify items that may be of interest to particular users. Recommender systems have become an integral part of our lives. People interact with recommender systems on a daily basis for leisure activities, such as finding a movie to watch, a friend to follow, etc., or for professional activities, such as finding a topic to study, finding co-authors to collaborate with, etc. Furthermore, online service providers rely heavily on recommender systems to increase their revenue, to diversify their item recommendations (and indirectly sales), and ultimately to keep their customers satisfied and engaged. Without recommender systems, it seems impossible for online businesses to survive in this competitive world-market. Most available recommender systems are focused on single-domain recommendation tasks (e.g., the task of recommending books to particular users, based on the book history of a large number of users). However, for users with a limited behavior history (i.e., a small number of interactions), single-domain recommender systems perform poorly, suffering from the data sparsity and cold start problems. Recent developments in the general area of transfer learning have led to Cross-Domain Recommender Systems, which exploit user behaviour information from other source domains and transfer this information to a target domain. This knowledge transfer helps alleviate the sparsity issue and produce more accurate recommendations in the target domain. This is possible because users generally interact with multiple, related domains in their daily lives. In this dissertation, I focus on cross-domain recommender system approaches, which employ state-of-the-art deep neural networks, to tackle the data sparsity problem in a target domain. The source domains used to gain additional information in the cross-domain setting are selected to have user overlap with the target domain. Specifically, I propose three novel cross-domain recommendation models, which leverage state-of-the-art single-task recommendation models. The first proposed cross-domain model is a non-sequential recommender system (i.e., the order of the items in a user history is not considered), and it uses a neural collaborative filtering approach to performs knowledge transfer between the source and target at the embedding layer level. The second proposed approach is focused on sequential recommendations, and uses the successful self-attention mechanism to identify useful item preferences in both source and target domains. It also uses the early fusion technique to combine a pre-learned global source representation of a user with a target representation of the user. Finally, the third proposed cross-domain model uses one or more source domains to obtain users' general preferences, while the target domain is used to extract users' current preferences. The two types of preferences, general and target-specific preferences expressed as representational vectors, are then fused together to achieve knowledge transfer across domains and improve the recommendation accuracy in the target domain, especially when the target domain faces the sparsity problem. Public cross-domain datasets are employed to evaluate our proposed models in different settings relevant to target sparsity. Experimental results show that our proposed models increase the recommendation accuracy in the target domain, outperforming existing state-of-the-art recommender system models