Dual Branch Graph Representation Learning-Based Approach for Next Point-of-Interest Recommendation

Next Point-of-Interest (POI) recommendation, a sub-task of POI recommendation, focuses on predicting the next POI a user will visit, relying on the user’s sequential check-in history. In this paper, we observe that existing methods for this task have a fundamental limitation: they find it difficult to comprehensively model the associations between POIs from both explicit and implicit perspectives. To bridge this gap, we present a new method DBGR. Specifically, DBGR first introduces a language model to extract semantic representations from the text-based features of POIs, such as category information. Subsequently, it constructs a semantic association graph which preserves the semantic relations between POIs and are further fed into a graph neural network-based backbone to learn the representations of POIs in the semantic feature space. Complementarily, based the global and local check-in patterns, the Transformer model is deployed to extract the representations of POIs and users from historical check-in records. Extensive results on real-world datasets have showcased the effectiveness of DBGR for next POI recommendation, compared to mainstream approaches.