STaTRL: Spatial-temporal and text representation learning for POI recommendation

With the rapid development of location-based social networks (LBSNs), point-of-interest (POI) recommendations have become a practical problem attracting more and more attention. Recent studies mostly utilize contextual features and sequential patterns of users’ check-ins to recommend POIs. However, there are still many deficiencies in existing works, such as: (1) insufficiently learning relations among far-apart visits in user check-ins; (2) not effectively incorporating geographical information when modeling user-POI interactions; and (3) little exploiting the features from reviews for the POI recommendation task. To tackle the above problems, we propose spatial-temporal and text representation learning (STaTRL), which employs Transformer to learn long-term dependencies among visits in the check-ins sequence and adopts an improved approach to compute the attention between visits by applying geographical information to the self-attention layer in Transformer. Meanwhile, users’ perspectives and POIs’ reputations learned from textual reviews are explored to improve the performance. In addition, a multi-task objective framework is adopted to simultaneously train the hidden representations of users’ historical check-ins trajectories which are shared by these two tasks. Concretely, STaTRL consists of (1) the principal task, i.e., unvisited POI recommendation that recommends to users the unvisited POIs, and (2) the auxiliary task, i.e., user’s POI preference learning whose candidates include both visited and unvisited POIs. We found that the latter task helped train the embedding of visited POIs and further boosted the performance of the former task, and lacking any of both would decline the performance. Extensive experiments on three public datasets demonstrated that STaTRL vastly outperformed the state-of-the-art methods.