Adaptive Fusion of Global Information for Position Aware Multi-Interest Sequential Recommendation

Sequential recommendation based on multi-interest networks has become a research hotspot cause it meets the various needs and preferences of users. However, current models only focus on how to extract multiple interests of users, but do not distinguish the importance of interests; and the global information that can simulate the evolution trend of user interest is not fully utilized. In this work, we propose a novel approach named adaptive fusion of global information for position aware multi-interest sequential recommendation(AGPM). Specifically, all users' historical interaction sequences are utilized to simulate the potential evolution trend of user interests. In this process, we set neighbor window size and use the position interval size between pairs of co-occurring items to adjust the weight between them, and design an adaptive strategy to fuse high- and low-order global information, enriching with highly relevant items while expanding high-order correlations. In addition, relative position interval information is used as the key factor to capture users' multiple interests, so as to highlight the content that users are more concerned about recently. Extensive experiments and analyses are carried out on three public datasets, and the experimental results demonstrate the effectiveness and superiority of the proposed method.