Global spatio-temporal aware graph neural network for next point-of-interest recommendation

Next Point-of-Interest (POI) recommendation is becoming increasingly popular with the rapidly growing of Location-based Social Networks (LBSNs). Most existing models only focus on exploring the local spatio-temporal relationships between POIs based on the trajectory sequence of current user. However, we argue that there exits not only local spatio-temporal relationship but also global spatio-temporal relationship, where two POIs are correlated if they appear in the trajectories of all users within certain geographical distance and time intervals. In this paper, we propose Global Spatio-Temporal Aware Graph Neural Network (GSTA-GNN), a model that captures and utilizes the global spatio-temporal relationships from the global view across the trajectories of all users. Specifically, we first break down the independence between trajectories and link all pairs of POIs based on the POI transitions to construct a global spatial graph and a global temporal graph. Then graph neural network is utilized to learn the global general representations of POIs. In addition, we introduce the spatio-temporal weight matrix, which converts the spatial and temporal intervals into suitable weight values and combines them in an adaptive manner. Then we propose to incorporate the spatio-temporal weight matrix into self-attention module of a multi-head self-attention layer to enrich the personalized representation of user trajectory. Experiments on three real datasets show that GSTA-GNN is superior to the state-of-the-art models in next POI recommendation task.