Deep Potential Geo-Social Relationship Mining for Point-of-Interest Recommendation

Point-of-Interest (POI) recommendation has been an important research topic in data mining and the popularity of location-based social networks (LBSNs) has significantly contributed to POI recommendation. The existing POI recommendation models mostly adopt various explicit social relationships under geographical space. The implicit relationships among users under a certain geographical region are rarely taken into account, though they have major influence on user behaviors. Due to the above limitations, we were motivated to introduce an innovative Deep Potential Geo-Social Relationship mining model for POI Recommendation (DPGSR-PR). The proposed DPGSR-PR performs the personalization of geographical features, towards the determination of user check-in behaviors and choices in specific domains, which is achieved by estimating and considering kernel density. Moreover, user preferences and personalized geo-social influence are incorporated into a geo-social recommendation framework under a holistic view. Specifically, the deep potential geo-social relationships include the explicit-implicit user geosocial relationships between users (EIU-GSR) and the implicit common check-in POI-based geo-social relationships (ICP-GSR). The estimation of Kernel density and the two-hop random walk approach are employed in an effort to mine the EIU-GSR. The ICP-GSR is discovered by specifically determining and considering the Jaccard similarity coefficient and kernel density estimation. Due to the fact that the role of both EIU-GSR and ICP-GSR as regularization terms is quite significant, we used their combined impact to obtain a unique recommendation model that employs matrix factorization. The proposed DPGSR-PR was tested on two datasets, which has proven that DPGSR-PR outperforms other well-established models.