Valuable Interactions, Valuable Recommendations: A New Approach for Integrating General User Preferences in Session-Based Recommender Systems

The advent of session-based recommender systems (SBRS) has tremendously contributed to giving recommendations to users without considering their historical data. These systems operate by recommending items based solely on the clicks (i.e., user-item interaction) within the particular session at hand, which represent short-term user preferences. Still, researchers have also attempted to integrate long-term user preferences into the process of SBRS. Recent studies have utilised graph neural networks (GNN) to achieve this goal. By feeding all session graphs into the GNN, we can extract the long-term user preferences by combining all item vectors from that session and regard the final interaction in the session as the user’s current interest (as current short-term preferences). We recognise that this method is somewhat constrained owing to the fact that the final session item may not fully represent the user’s current interest. For the purpose of better understanding user interests, this paper presents a new approach by introducing a voting strategy that progressively learns and determines the most relevant interaction within the session, which, in the best-case scenario, approximates user intent best. The embedding of this item is then processed in conjunction with the global embedding (all item vectors combined) and the last item embedding within an attention mechanism to integrate user’s general preferences in SBRS accurately. The experimental results on two well-known real-world datasets show the superiority of the proposed approach compared to the state-of-the-art methods in SBRS, especially when low-temperature values are set in the voting layer.