A study of neighbour selection strategies for POI recommendation in LBSNs

Abstract

Location-based Recommender Systems (LBRSs) are gaining importance with the proliferation of location-based services provided by mobile devices as well as user-generated content in social networks. Collaborative approaches for recommendation rely on the opinions of liked-minded people, so called neighbors, for prediction. Thus, an adequate selection of such neighbors becomes essential for achieving good prediction results. The aim of this work is to explore different strategies to select neighbors in the context of a collaborative filtering based recommender system for POI (places of interest) recommendations. Whereas standard methods are based on user similarity to delimit a neighborhood, in this work several strategies are proposed based on direct social relationships and geographical information extracted from Location-based Social Networks (LBSNs). The impact of the different strategies proposed has been evaluated and compared against the traditional collaborative filtering approach using a dataset from a popular network as Foursquare. In general terms, the proposed strategies for selecting neighbors based on the different elements available in a LBSN achieve better results than the traditional collaborative filtering approach. Our findings can be helpful both to researchers in the recommender systems area as well as to recommender systems developers in the context of LBSNs, since they can take into account our results to design and provide more effective services considering the huge amount of knowledge produced in LBSNs.