CPGMA: Spatiotemporal Heterogeneous Graph Convolution with Multi-scale Attention for Sequential POI Recommendation​

With the widespread adoption of mobile social networks, sequential Point-of-Interest (POI) recommendation has become a key technology for enhancing user experience and generating commercial value. This paper proposes the CPGMA model, which integrates Graph Convolutional Networks (GCN) with Multi-Scale Linear Attention, to effectively capture complex spatiotemporal correlations in user check-in behaviors by constructing a spatiotemporal heterogeneous graph. The model innovatively combines GCN with a multi-scale attention mechanism to enable joint modeling of users' long-term and short-term preferences. Experimental results demonstrate that CPGMA significantly outperforms existing state-of-the-art methods on multiple public datasets, especially in sequential POI recommendation scenarios, where it shows clear advantages.