Integrating Users’ Long- and Short-Term Preferences for Session-based Recommendation
Hanlin Liu, Zhuoming Xu, Qianqian Zhang, Yan Tang · 2022 IEEE 25th International Conference on Computer Supported Cooperative Work in Design (CSCWD) · 2022
Different from conventional recommendation models which typically capture long-term yet static user preferences, session-based recommendation (SR) models aim to capture short-term but dynamic user preferences to provide more timely recommendations. Recent studies have shown that incorporating users’ long-term preferences into the SR model can help improve recommendation accuracy. However, since information on the time intervals between interactions is not fully exploited and/or long-term and short-term preferences are not effectively distinguished, most existing SR models that explicitly/implicitly capture users’ long-term and short-term preferences have defects in preference mining and integration, which leads to inaccurate recommendations. This paper addresses the problem of recommending the next interaction (item) given a user’s current session and some historical sessions by effectively integrating the user’s long-term and short-term preferences. We propose a novel model named Integration of Long- and Short-Term Preferences (ILSTP) for session-based recommendation. ILSTP employs time interval embedding function, self-attention, and hybrid aggregator combining max pooling and average pooling to capture long-term preference from the historical sessions; it utilizes gated graph neural network to capture short-term preference from the current session; it integrates the captured long-term and short-term preferences through attention aggregator. Experiments show that ILSTP achieves new state-of-the-art performance in next interaction (item) recommendation.