A Graph-Enhanced Dual-Granularity Self-Attention Model for Next POI Recommendation

Abstract

The growth of location-based social network (LBSN) services has resulted in a demand for location-based recommendation services. The next point of interest (POI) recommendation is identified as a core service of LBSNs. It is designed to provide personalized POI suggestions by analyzing users’ historical check-in data. General methods model users’ check-in sequences by directly applying attention mechanisms. However, they often overlook the importance of global information from other users’ behaviors, and the embeddings of check-ins are not sufficiently effective. This approach fails to capture the collective influence of multiple check-ins. To address this issue, we propose a graph enhanced dual-granularity self-attention model (GEDGSA) that can model users’ preferences from both fine-grained and coarse-grained perspectives to improve prediction performance. First, a graph-enhanced embedding module is designed to capture common transition patterns among all users to obtain initial POI features. Second, the virtual trajectory construction operation is introduced to transform multiple check-ins into coarse-grained virtual check-in items. The GEDGSA learns user check-in sequences from both fine-grained and coarse-grained perspectives. Finally, our method experiments on the Foursquare-NYC and Foursquare-TKY datasets, demonstrating that it outperforms most existing methods.