Contextual Models for Sequential Recommendation

Recommender systems aim to capture the interests of users in order to provide them with tailored recommendations for items or services they might like. User interests are often unique and depend on many unobservable factors including internal moods or external events. This phenomenon creates a broad range of tasks for recommendation systems that are difficult to address altogether. Nevertheless, analyzing the historical activities of users sheds light on the characteristic traits of individual behaviors in order to enable qualified recommendations.

In this thesis, we deal with the problem of comprehending the interests of users, searching for pertinent items, and ranking them to recommend the most relevant items to the users given different contexts and situations. We focus on recommendation problems in sequential scenarios, where a series of past events influences the future decisions of users. These events are either the developed preferences of users over a long span of time or highly influenced by the zeitgeist and common trends. We are among the first to model recommendation systems in a sequential fashion via exploiting the short-term interests of users in session-based scenarios.

We leverage reinforcement learning techniques to capture underlying short- and long-term user interests in the absence of explicit feedback and develop novel contextual approaches for sequential recommendation systems. These approaches are designed to efficiently learn models for different types of recommendation tasks and are extended to continuous and multi-agent settings. All the proposed methods are empirically studied on large-scale real-world scenarios ranging from e-commerce to sport and demonstrate excellent performance in comparison to baseline approaches.