MODELING SPATIO-TEMPORAL DYNAMICS WITH TRANSFORMER ATTENTION FOR POINT OF INTEREST RECOMMENDATION

Abstract

Point-of-Interest (POI) recommendation is one of the most important tasks in location-based services to recommend individuals locations based on their past check-ins, spatial interests, and temporal patterns. In this work, we introduce a new method that learns spatio-temporal dynamics via Transformer-based attention to enhance recommendation precision. We represent user and POI IDs via embedding layers and bring geographical context in by normalizing and embedding latitude-longitude points. Temporal relationships between user check-in sequences are modelled with a Transformer Encoder to allow for parallel sequence modelling and learning of distant interactions. Temporal information is combined with user and spatial representations to provide a common latent feature space, which is then passed through a fully connected layer to provide POI probability scores. The model is trained with negative log-likelihood loss and optimizes Adam with gradient clipping for stability. Evaluation by Accuracy@k, Precision@k, Recall@k,F1@k and NDCG@k presents ranking performance of the proposed model on effective POIs. The proposed approach yields an interpretable and scalable solution to next-POI recommendation with deep consideration of spatial, temporal, and behaviour patterns collectively.