KEAN: Knowledge Embedded and Attention-based Network for POI Recommendation

Abstract

In recent years, with the rapid development of location-based social networks (LBSN) in the Internet, point of interest (POI) recommendation has become a hot spot. Most existing researches make use of contextual information to model users' interest preferences. However, the existing methods for extracting various auxiliary information still need to be improved, such as how to treat the users' social relations equally. In order to obtain users' actual preferences more accurately, in POI recommendation, we propose a deep learning framework KEAN (Knowledge Embedded and Attention Based Network) based on knowledge graph and attention model. The framework includes knowledge-graph embedding method, preference extraction network based on attention mechanism and recommendation network. Our study used knowledge-graph embedding method to get the embedding of each user and POI. In addition, an LSTM network based on attention mechanism was proposed, which uses LSTM network to learn the user's preferences according to the user's check-in sequence. Besides, the attention mechanism was used to extract friends' preferences and merge them with the user's preferences to generate end-user preferences. Finally, our model use fully-connected neural networks to realize recommendations. The effectiveness of the model was proved by the experimental results based on real LBSN datasets.