Multivariate Hawkes Spatio-Temporal Point Process with attention for point of interest recommendation

Xin Zhang, He Weng, Yuxin Wei, Dongjing Wang, Jia Chen, Tingting Liang, Yuyu Yin

Abstract: Point-of-interest (POI) recommendation is one of the most essential services in modern Location-Based Social Networks (LBSNs) to address the problem of information overload presented in various online applications. However, existing methods may ignore the spatio-temporal context, especially the interplay between time and space in user behavior patterns, and treat them as independent features. To fully exploit the spatio-temporal information for accurate POI recommendations, we propose a Multivariate Hawkes Spatio-Temporal Point Process with attention (MHSTPP-a), which predicts the user’s preference for the next POI by jointly modeling the correlations between users and POIs. In particular, a novel and lightweight representation learning approach is designed to model user interaction sequences from the perspective of spatio-temporal point processes. Specifically, MHSTPP-a first learn the informative embeddings of users and POIs based on historical check-in records. Then, a multivariate Hawkes process with the attention mechanism is designed to model the spatio-temporal dependencies and correlations between check-in behaviors in continuous time and space and learn users’ general and contextual preferences. Finally, MHSTPP-a can perform accurate next POI recommendations according to personal preferences. Extensive experiments carried out on three real-world datasets demonstrate that the proposed approach MHSTPP-a outperforms baselines, including several state-of-the-art methods in various evaluation metrics.