Abstract: Friendship prediction has become one of the major studies of location based social network (LBSN)．This paper proposed an approach for predicting friendship,which fuses the topology network and geographical features of LBSN．We first adopted the information gain to measure the contribution of different features to human friendship,and chose three key features:user social topology,the category of the location where people check in,and check in points．We then presented the friendship prediction method based on the fusion of the selected features．Three different classification models,including Random Forests,Support Vector Machine (SVM),and Naive Bayes,were selected to predict human friendship．Experimental results on the real collected data from Foursquare and JiePang verify the efficacy of the selected features and the accuracy of friendship prediction.

Key words: Location-based social network,Friendship prediction,Information gain,Feature fusion