Research On Point-of-interest Recommendation Algorithms Based On Location-based Social Network

Point-of-Interest(POI)recommendation,which focuses on helping each individual user to discover her potential preference locations,has become an important research problem since the location-based social networks(LBSNs)has been developing rapidly in recent years.Different from other recommendation problems,POI recommendation can utilize rich contexts such as user's check-in records,geographical and categorical information of locations,and so on.Among these contexts,the categorical information of checked-in locations implicitly indicates the topic of a user's activity at a time,which is useful to depict the user's individual check-in preference.However,previous studies usually lack of fully utilizing the categorical information.For example,they mainly model the categorical information as a preference bias separated from other factors rather than use it to enhance the other factors.To solve this problem,in this paper we propose a novel POI recommendation model,namely Geography-based and Category-Enhanced Matrix Factorization(GeoCaMF),which exploit categorical information to enhance the User-based Collaborative Filtering(UCF)technique and Itembased Collaborative Filtering(ICF)technique while incorporating them into weighted matrix factorization(WMF).Specifically,we firstly cluster users into groups according to their check-in frequency distribution on location categories to gather users with high similarity of check-in preference,which can help enhance the user similarity calculation while applying UCF for each user within these groups.Then,we introduce the categorical information to enhanced the geographical-distance-based location similarity while modeling the geographical neighborhood influence for each location by applying ICF.Finally,we integrate UCF and ICF into WMF model to learn the potential preference of locations for each individual user.Performance evaluations are conducted for our proposed model by comparing with several state-of-the-art methods on two large-scale LBSNs datasets under four metrics including Precision@K,Recall@K,nDCG@K and MAP@K.The experiment result demonstrates the effectiveness of our model.