Towards recommendation with user action sequences

Across the web and mobile applications, recommender systems are relied upon to surface the right items to users at the right time. This implies user preferences are usually dynamic in real-world recommender systems, and a user's historical action records are not equally important when predicting her/his future preferences. Most existing recommendation algorithms, including both shallow and deep approaches, usually treat all user's historical actions equally, which may have lost order information between actions. In this thesis, we study the problem of modeling user action sequences for recommendation~(a.k.a sequential recommendation). Motivated by the distinct challenges when modeling user sequences, we focus on building sequential recommendation models to capture various types of dependencies (sequential patterns). In particular, the dependencies can be in different forms. Also, they can either from the local part or the long-tail of user sequences. Though usually neglected in existing approaches, these dependencies are informative for accurate prediction of user preference. In our first work, we discover the dependencies in real user sequences can have two different forms: point-level and union-level. We propose a unified model to jointly capture both forms of sequential patterns. In our next work, we analyze the property of dependency from different temporal ranges of long user sequences. Based on our observation, we propose a neural mixture model as a tailored solution to deal with dependencies from all temporal ranges. Finally, inference efficiency is critical for each model since recommendation is an online service. It is particularly important for sequential recommendation as user's sequence frequently changes and inference is needed with the new sequence. We provide a knowledge transfer framework to satisfy the efficiency requirement for recommendation models. We show this framework can be used to learn a compact recommendation model with better inference efficiency but with the similar efficacy of a large model. Our proposed solution can be also used for other ranking problems.