Memory-Enhanced Gated Graph Attention Network for Next Point-of-Interest Recommendation

Abstract

The recommendation of the next Point-of-Interest (POI) has become an essential component of location-based social networks (LBSNs). Its purpose is to advise users' undiscovered POIs based on their check-in record. However, the existing work mainly regards the long-term interest of users as a periodic and stable change, but ignores the dynamic change. Besides, sparse POI-POI transformation limits the ability of the model to effectively learn contextual information in local views. We present a memory-enhanced gated graph attention network (MEGGA) in this paper. In order to effectively capture context in a local view, a gated graph attention network (GGAN) is designed, where category information is introduced to alleviate sparsity. In the long-term module, combined with the advantages of RNN and memory network, users' interests are dynamically captured by constantly updating user sequence while processing sequence information. Moreover, we use the attention layer to integrate multiple interests and learn the degree of users depending ¹ on different interests. Two public datasets were used in experiments to confirm MEGGA efficacy.