Session-Based Recommendation with GNN and Time-Aware Memory Network

The goal of the session-based recommendation system (SBRS) is to predict user’s next behavior based on anonymous sessions. Since long-term historical information users not available, deep learning technology has become mainstream in systems instead traditional content-based methods. However, most SBRS methods only consider session itself, ignoring collaborative from other Even if some models collaborations between sessions, they mostly use click order calculate similarity and ignore time user spends different items, which might imply varying interest these items. In this paper, we propose a model with GNN time-aware memory networks (SR-GTM), learns representation by combining itself relevant neighbor Specifically, SR-GTM mainly includes inner feature extraction module (IFEM) outer (OFEM). IFEM uses learn features its item sequence, OFEM network dwell encoded extract information. Finally, aggregates gating mechanism then decodes output softmax layer obtain score for each candidate item. Experiments three public datasets Yoochoose1/64, Yoochoose1/4, RetailRocket show that achieves optimal performance compared state-of-the-art More specifically, improvements 0.77%, 0.38%, 3.63% over best baseline method P@20 2.91%, 2.52%, 2.49% MRR@20, respectively.