PAPR:Periodic Aware Spatial temporal Network for POI Recommendation

The next POI recommendation plays a significant role in location-based services because it provides personalized suggestions for destinations to users. The most advanced research utilizes enhanced attention mechanisms or neural graph networks to handle temporal and spatial factors. However, existing methods mainly concentrate on simulating users’ complex transition patterns while neglecting the geographical proximity and complex periodicity of the POIs users visit. In this paper, a transformer-based model is proposed for the next POI recommendation in order to figure out periodic intervals. The multi-head attention mechanism of the transformer is used to capture the intricate interdependencies between POIs in the user check-in sequence. The periodic interval matrix is added to the attention mechanism as an inductive bias, allowing the multi-head attention mechanism to analyze the systematic relationships between all POIs under the periodicity constraint. In addition, as an auxiliary job, we offer a distance loss function that reflects the geographical distance between the predicted POI and the actual visited POI by the user. The distance loss function assures that the predicted and natural geographical spaces are consistent. The experimental results based on two classical LBSN datasets demonstrate the excellent performance of PAPR, which improves by 2.04% and 4.30% relatively compared to the state-of-the-art model.