LBSN-based Personalized Routes Recommendation

Recently, online social network on internet and user’s activities in offlinephysical world have been studied very well, so the study of heterogeneousnetwork which combines online social network and offline physical worldbecomes meaningful, such as the more and more popular Location Based SocialNetwork (LBSN), it is a new platform to understand user’s behavior andpreference through location information.In this paper, we focus on Location Based Social Network. From the aspectof regularity of user’s movement and describe user’s activity regularity onsemantic space, i.e. location’s category information, to recommend personalizedroutes to a certain user which match his/her activity regularity.We consider location’s semantic meaning, as use both location’sgeographical information and semantic information can better describe user’smovement, so we process user’s raw trajectory and extract location trajectorieson geographical space and category trajectories on semantic space to describeuser’s movement, and get user’s experiences under different location categories,to build user model. Mine personalized information to describe user’s activityregularity on category trajectories, i.e. activity pattern in our paper, according touser model to calculate user similarity, extract candidate routes which matchuser’s activity pattern from high similarity users, through scoring strategy whichconsider the information of user’s own activity pattern, user whole similarity,user similarity under different categories and location’s popularity under its owncategory, to recommend top-k score routes to user. We define activity pattern onsemantic space, compare to traditional methods, this method considers location’ssemantic meaning can also work even if the area is strange to users.We use GeoLife dataset and POI dataset which contain60%POIs in Beijingas our experimental data. Experimental results show that about35%users haveactivity pattern, so personalized routes recommendation is meaningful. Ourrecommendation method is better than traditional cosine based method onprecision, K-cover and recall.