Efficient point-of-interest recommendation with hierarchical attention mechanism

Abstract Personalized Point-of-Interest (POI) Recommendation is very important for application platforms based on Location Based Social Networks (LBSNs). It can assist users in making decisions to alleviate the problem of information overload, and can also improve the user experience of these platforms and advance platform operators achieve personalized and accurate advertising. However, there exist some problems of data sparseness and cold start for a single user, and it is also difficult to mine valuable long-tailed POIs, although the size of the check-in data is large. Therefore, in order to address the above problems, we propose a personalized POI Recommendation approach based on Hierarchical Attention Mechanism (HAM-POIRec) which can effectively increase data utilization. Firstly, we define the concepts of explicit features and implicit features, which pave the ideas of selecting data and computational models for POI recommendation based on machine learning. Secondly, we propose a hierarchical attention mechanism with the structure of local-to-global, which extracts contributions and mines more hidden information from individual features, combination features, and overall features. Finally, we present the Natural Language Processing (NLP)-based “User-POI” matching mechanism for the first time in the field of POI recommendation to improve the recommendation accuracy by fine-tuning the POIs predicted by the recommendation system. Extensive experiments are conducted for demonstrating that the HAM-POIRec method outperforms state-of-the-art DeepPIM method and the other comparison methods (SAE-NAD, MGMPFM and LRT), especially in predicting sequence POIs and solving cold start problem.