STMG: Spatial-Temporal Mobility Graph for Location Prediction

Location-Based Social Networks (LBSNs) data reflects a large amount of user mobility patterns. So it is possible to infer users' unvisited Points of Interest (POIs) through the users' check-in records in LBSNs. Existing location prediction approaches typically regard user check-ins as sequences, while they ignore the spatial and temporal correlations between non-adjacent records. Moreover, the serialized form is insufficient to analog user complex POI moving behaviors. In this paper, we model user check-in records as a graph, named Spatial-Temporal Mobility Graph (STMG), where the nodes and edges fuse the spatial-temporal information in absolute and relative aspect respectively. Based on STMG, we propose a location prediction model named Spatial-temporal Enhanced Graph Neural Network (SEGN). In SEGN, the STMG nodes are encoded as the embeddings with specific time and location semantics. Last but not the least, we introduce three kinds of matrices, which completely depict the user moving behaviors among POIs, as well as the relative relationships of time and location on STMG edges. Extensive experiments on three real-world LBSNs datasets demonstrate that with specific time information, SEGN outperforms seven state-of-the-art approaches on four metrics.