MODELING DYNAMIC ASPECTS OF CONTEXT-AWARE RECOMMENDER SYSTEMS

Abstract

We focus on how to effectively design context-aware recommender systems in non-stationary environments. Firstly, we present two strategies that are based on Self-Paced Learning and the rating profiles of users, items, and contextual conditions to select discriminative instances (i.e., user interactions) from a stream of incoming instances. We aim to make the online updating process efficient via selecting the best instances while reducing the impact of information loss. Secondly, we present how to efficiently infer ever-changing latent contexts using a dynamic network embedding method and integrate these contexts with an incremental recommender system. We focus on the most discriminative nodes and design a biased random walk to explore the nodes that are changed over time, with higher transition probabilities. Finally, we show how to use the classical Multi-armed bandits with Matrix Factorization by modeling the changes in user curiosity to enhance accuracy and temporal diversity.