POI Recommendation Based on Graph Enhanced Attention GNN

Abstract

With the continuous progress of mobile Internet technology and GPS positioning technology of mobile devices, Social Network and Location Based Services (LBS) are gradually converging to form Location Based Social Network (LBSN). POI (Point of Interest) recommendation systems face the problems of variable user interests, very sparse user and POI check-in matrices, and nonlinear interaction modeling. To address the above problems, a Graph-enhanced Attention Graph Neural Network model is proposed for POI recommendation (POI-GAGN in short). POI-GAGN mines user/POI node representations on user-POI interaction graph, user-user social interaction graph, and POI-POI association interaction graph through interaction node feature extraction module, learns POI attribute information representations through text feature extraction module, and extracts short-term preference representations of users through short-term preference extraction module. A graph-enhanced attention mechanism is designed to correlates node representations, attribute information representations of POI, and short-term preferences of users with each other to achieve better information fusion. Finally, we conduct sufficient experiments on two real datasets to prove that the recommendation effect of POI-GAGN is better than other current advanced POI recommendation methods, and POI-GAGN can better overcome the problems of data sparsity and cold start in recommendations.