Temporal Dynamics in Recommender Systems

In real-world scenarios, user preferences for items are constantly drifting over time as item perception and popularity are changing when new fashions or products emerge. The ability to model the tendency of both user preferences and item attractiveness is thus vital to the design of recommender systems (RSs). However, conventional methods in RSs are incapable of modeling such a tendency accordingly, leading to an unsatisfactory recommendation performance. This thesis proposes a framework for the temporal dynamics problem in RSs. The temporal properties and dynamic information in user preferences and item attractiveness derived from user feedback over items are modeled, learned and applied to predict user preferences on items over time. The framework provides original solutions to improve the performance of RSs by incorporating and exploiting this significant but traditionally neglected information. Firstly, a novel probabilistic temporal model for RSs is developed to tackle the inherent nonlinear and non-Gaussian dynamic problem with the complex and diverse real-world recommendation scenarios. It tracks simultaneously latent factors that represent user preferences and item attractiveness. A learning and inference algorithm combining a sequential Monte Carlo method and the Expectation-Maximization algorithm for this model is developed to tackle the top-k recommendation problem over time. Secondly, a novel probabilistic personalized and item-wise temporal model is proposed to solve the cold start transition (CST) problem by collaborative tendencies without any prior assumptions about the structure of the dynamics. The CST problem is first defined in this thesis, which is a result that users often leave feedback on an item only once and on only one period, preventing from learning any dynamics directly. Finally, a Bayesian Wishart matrix factorization method is proposed to model the temporal dynamics of variances due to sudden changes and other local temporal effects among user preferences and item attractiveness. It combines the collapsed Gibbs sampling method and the elliptical slice sampling method.The presented models and learning algorithms are validated experimentally on several real-world public benchmark datasets. The experimental results demonstrate that those models and algorithms significantly outperform a variety of state-of-art methods in RSs