Multi-features Point-of-Interest (POI) recommendation using graph neural network

Access paper here

Point-of-Interest (POI) recommendation is a progressively emerging field that suggests unvisited locations to users based on their check-in history. To overcome the cold-start and data sparsity problems, numerous solutions have incorporated combination of spatial, temporal, social, and textual features. Additionally, recently proposed methods have leveraged graph neural networks, which have shown a significant improvement in recommendation performance. However, existing solutions have considered spatial features in conjunction with social, temporal, or textual aspects. Moreover, they have only leveraged recommendations based on users' check-in history, ignoring other factors. For example, if a user checks in at a park but likes the food, it means she is more interested in the food than in the park itself. Considering textual features can provide more personalized recommendations as compared to exploiting only users' check-in history. In this regard, exploiting textual reviews provided by users related to the desired locations can help achieve this goal. In literature, relatively few existing solutions have implicitly incorporated combination of these features; however, they failed to explicitly combine them all in a single solution. Furthermore, they have considered the embedding of User-User and user-POI interaction networks, while plenty of studies ignored the explicit embedding of the POI-POI network, which can effectively capture spatial influence. Therefore, the proposed study has extended the functionality of DiffNet++, a graph-based approach that considers social and user-item interaction embedding, incorporating high-order node proximity. Also incorporated spatial, temporal, social, and textual features into a single solution, and included a location network to effectively leverage spatial influence in the recommendations. Moreover, we have successfully established high-order node proximity within User-User, User-Point of Interest (POI), and POI-POI networks. To ensure the effectiveness of the proposed method, we utilized the Yelp dataset, which has been evaluated based on precision, recall, hit ratio, and normalized discounted cumulative gain. Based on the experimental results, the proposed study outperformed the baselines in terms of performance.