Understanding dynamic user preferences using user-centric reservoir model for enhancing personalization

Abstract

Streaming Session-based Recommender Systems (SSRS) generate personalized recommendations from continuously arriving session data in the form of clickstreams. However, the massive arrival of clickstreams causes the forgetting of learned user preferences, necessitating periodic retraining of the SSRS model. This requires storing clickstreams in the reservoir; however, its limited storage space causes existing approaches to randomly replace and sample clickstreams, leading to the loss of user preferences that are unique compared to other users. Moreover, retraining the SSRS model with all the clickstreams in the reservoir results in an overload scenario. This work introduces a user-based partitioning approach that allocates dedicated reservoir partitions for individual user interactions. Furthermore, a priority-based sampling approach is introduced to select clickstreams from each partition for retraining. By considering both historical and ongoing user preferences, the proposed methodology generates personalized recommendations for the users. Experimental results on benchmark datasets demonstrate its superiority over state-of-the-art baseline methodologies. © Bharati Vidyapeeth's Institute of Computer Applications and Management 2025.