A Time Interval Aware Approach for Session-Based Social Recommendation

Abstract

Users on social media like Facebook and Twitter are influenced by their friends. Social recommendation exploits users’ social activities for modeling their interests to enhance the activeness and retention of users. Besides, their interests change from time to time. Session-based recommendation divides users’ interaction history into sessions and predict users’ behaviors with the context information in each session. It’s essential but challenging to model the social activities and the dynamic property in an unified model. Besides, most of existing session-based recommendation approaches model users’ interaction history as ordered sequence in regardless of real timestamps of those interactions. To solve the above issues together, we design a heterogeneous graph for modeling the complex interactions among users and items and propose a Time Interval aware graph neural network-based Recommendation approach(TiRec) to model both the social activities and the dynamic property of users’ interaction with items. Furthermore, to capture users’ dynamic preference, we propose a time interval aware and self-attention based aggregator to model users’ preference in each session. Experimental results on several real-world datasets demonstrates the effectiveness of our proposed approach over some competitive baselines.