Self-training Temporal Dynamic Collaborative Filtering

Abstract

Recommender systems (RS) based on collaborative filtering (CF) is traditionally incapable of modeling the often non-linear and non Gaussian tendency of user taste and product attractiveness leading to unsatisfied performance. Particle filtering, as a dynamic modeling method, enables tracking of such tendency. However, data are often extremely sparse in real-world RS under temporal context, resulting in less reliable tracking. Approaches to such problem seek additional information or impute all or most missing data to reduce sparsity, which then causes scalability problems for particle filtering. In this paper, we develop a novel semi-supervised method to simultaneously solve the problems of data sparsity and scalability in a particle filtering based dynamic recommender system. Specifically, it exploits the self-training principle to dynamically construct observations based on current prediction distributions. The proposed method is evaluated on two public benchmark datasets, showing significant improvement over a variety of existing methods for top-k recommendation in both accuracy and scalability.