Research On Friend Recommender Systems Based On Check-in Data In LBSNs

Nowadays,more and more people use their smartphones to access mobile Internet.In the meanwhile,they use various social networks to communicate with their each other,acquire information and share contents.Among these social networks,location-based online social networks(LBSNs),which utilize localization technique of smartphones,enable users to get their locations in real time.By using LBSNs,people can share their locations by checking in at some spots,find interesting places and make friends.With the popularity of LBSNs,more and more social networks allow users to attach their location information to contents they share.As a result,these social networks hold great amount of check-in data of users.These data contain much information of users' interests,which can be used to make various recommendations to users.Existing studies about friend recommendations in LBSNs focus on mining users' interests from position and semantic information of check-in data,and they often validate their recommender systems on location data collected from volunteers.In light of above deficiencies,we make full use of time information in check-in data and propose two friend recommender systems in LBSNs: semantic and time based friend recommender system and time relative entropy based friend recommender system.The main contents of this paper are as follows:1.This paper proposes a semantic and time based friend recommender system.In this system,to overcome the sparsity of data,we use the semantic information of check-in spots to classify the check-in data.Then we use TF-IDF(term frequency-inverse document frequency)model to achieve a balance between users' interests and popularity of locations.At last,we use relative entropy of check-in time among users to adjust cosine similarity and make friend recommendations.2.In order to improve the precision,recall and F-measure of recommender system above,we propose a time relative entropy based friend recommender system.We take check-in time as primary consideration in friend recommendations.After calculating relative entropy of check-in time among users,we picked users' most interested location categories using a TF-IDF based method.At last,we take users with least relative entropy in these categories as friend recommendations.3.We evaluate our friend recommender systems on check-in data collected from Gowalla,a famous location-based social network.When running the time relative entropy based friend recommender system,we use supervised learning methods in machine learning.The experiment results show that these two friend recommender systems outperform classic collaborative filtering recommender system.