Enhancing spatio-temporal semantics with contrastive learning for Next POI recommendation

Next Point-of-Interest (POI) recommendation offers significant value for both location-based service providers and users. Existing next POI recommendation models usually rely on a supervised paradigm, using observed user-POI interactions to learn model parameters and data representations for final POI prediction. However, these models suffer from sparse supervised signal issue. Meanwhile, the overemphasis on the final POI recommendation performance results in insufficient representation of the spatio-temporal semantics in check-in sequences. To this end, we propose a model enhance Spatio-Temporal Semantics with Contrastive Learning (STSCL) for next POI recommendation, which effectively capture the semantics in the user sequential behavior. The main idea of our approach is to utilize the spatio-temporal semantics to create contrasting views via pre-training methods for improving recommender performance. Specifically, time and distance intervals are considered to divide the user’s entire sequence into more coherent subsequences. We design three contrastive learning objectives to learn the correlations among POIs, subsequences, and sequential transitions by utilizing the semantics of spatial, temporal, and context, respectively. Extensive experiments on three real-world datasets show that STSCL significantly improves next POI recommendation performance. The source code is available at: https://anonymous.4open.science/r/STSCL.