TOWARDS COMPREHENSIVE USER PREFERENCE LEARNING: MODELING USER PREFERENCE DYNAMICS ACROSS SOCIAL NETWORKS FOR RECOMMENDATIONS

Abstract

Recommendation Systems (RSs) continue to suffer from incomprehensive user preference learning, due to two primary reasons. First, incomplete user profiles of new and existing users lead to cold start and data sparsity problems. Second, ever-changing user preferences make recommendations obsolete over time. Therefore, we introduce time aware cross-network RSs that use auxiliary user preference information from multiple social networks to generate timely and holistic user preference profiles for recommendations. First, we show the feasibility and effectiveness of our approach using linear and deep learning-based solutions. Second, we show its practicality using deep learning models that are incrementally updated with oncoming user interactions across networks. Third, we introduce the new task of synthetically generating missing user preferences for users whose accounts on other networks are unknown. Experimental results showcase that the proposed solutions achieve superior recommendation quality and they consistently outperform state-of-the-art RSs in terms of recommendation accuracy, novelty and diversity.