Research On Classification Techniques In Location-Based Social Networks

With the popularization of mobile communication networks and mobile intelligent terminals,Location-Based Social Network(LBSN)services have had an advance rapidly development.Because more and more users make check-in activities,there is a large amount of users' check-in data in LBSNs.It brings the change and convenience to people's life and business by mining LBSN data that contains the information about people's interests,behaviors,patterns of life,etc.Classification is the core and basic technology of LBSN data mining,and it is one of the important technical means to solve the relevant problems in LBSNs,such as point-of-interest(POI)semantic annotation problem,recommendation problem,etc.However,in the face of the unique characteristics of LBSN data,the existing classification techniques still have some shortcomings.First,with the continuous increase of users,check-ins and information sharing in LBSNs,LBSN data increases sharply.In the face of a large amount of LBSN data,the efficiency of existing classification algorithms is significantly reduced.Second,LBSN data not only has the social relationship attribute,text attribute,time attribute of traditional social networks,but also unique the geo-spatial attribute.However,the existing classification techniques do not make full use of the rich attributes provided by LBSN data,which leads to the low classification accuracy of the algorithms.Therefore,the classification technology in LBSNs needs to be further researched.Based on the related problems in LBSNs,this paper makes full use of the rich attributes provided by LBSN data,and combines the Extreme Learning Machine(ELM)classifier with extremely fast learning speed to propose a series of high-performance classification methods.The contributions of this paper are summarized as follows:(1)POI semantic classification method based on the similar user.Aiming at the phenomenon of semantic missing on some POIs,this paper studies the semantic annotation of POIs.The semantic annotation of POIs can be used as the multi-label classification problem in LBSNs.First,similar users are found by capturing the similarity of different users'check-in activities in LBSNs.And POI semantic feature SUP based on similar users is proposed.Then,POI semantic classification method MSA-ELM is proposed based on SUP feature.In MSA-ELM method,an ELM classifier is trained for each tag in the tag space to support the multi-label classification.Finally,a series of experiments are designed to verify the effectiveness of SUP feature and the performance of MSA-ELM method.(2)POI classification method based on the check-in frequency.The classification problem is defined according to the prediction of the user's future check-in frequency on the POI that the user has never visited in LBSNs.First,POIC-ELM classification method is proposed.In POIC-ELM method,nine features are extracted from three aspects in LBSNs,including POI itself,individual user and social relationship.Then these features are used to train ELM classifier.Then,for the features extracted by POIC-ELM method,FS-ELM feature selection method is proposed based on ELM.Finally,a series of experiments are designed to verify the performance of POIC-ELM and FS-ELM methods.(3)ELM classification method to POI recommendation.POI recommendation in LBSNs is studied as a classification problem.First,for POI recommendation to individual users,PR-ELM classification method is proposed based on ELM.In PR-ELM method,four features are extracted from LBSN attributes,including POI popularity,distance from the user to POI,user's interest preference and social relationship.Next,these features are used to train ELM classifier.Then,for POI recommendation to a group of users,PGR-ELM classification method is proposed based on ELM.Different fromPR-ELM method,in PGR-ELM method,a new POI group recommendation feature is extracted by combining the interest preference of group members with the intimacy of group members.Finally,a series of experiments are designed to verify the performance of PR-ELM and PGR-ELM.(4)ELM classification method to friend recommendation.Friend recommendation in LBSNs is studied as a classification problem First,FE-ELM classification method is proposed based on ELM.In FE-ELM method,three corresponding features are proposed based on the spatial-temporal attribute,social attribute and text attribute in LBSNs,which ensures the classification accuracy.Next,these features are used to train ELM classifier.ELM with a fast training speed ensures the classification efficiency.Then,FE-DELM classification method is proposed based on the distributed ELM.Finally,a series of experiments are designed to verify the performance of FE-ELM and FE-DELM methods.