Mobile Users' Location Data Association Analysis And Anomaly Detection Technology

The research on technologies of mobile users' location data association analysis and anomaly detection which has important theoretical value and wide application prospect is needed urgently in the field of business recommendation,urban computing,social management and public safety.Focusing on mobile users' location data,four kinds of technologies for association analysis and anomaly detection were studied in this thesis,including mobile users' identity matching,social relationship type inference,group discovery and abnormal behavior detection.(1)User identity matchingTo solve the problems in user identity matching,such as solidification on user set,singleness of data source,difficulties in obtaining background information and so on,a new method for mobile users of heterologous location data was proposed.The method depended on the user's social relationships and behavior patterns in different location datasets.Firstly,location data was transformed into user's social relationship,personal location and spatio-temporal co-occurrence area(STCOA)information.Then active users were extracted according to social relationships and STCOA,and input matching algorithm as users to be matched.The algorithm which matched user identity was used to calculate the social relationship structure similarity and behavior pattern similarity of the users to be matched.Finally,according to the social relationship structure,the neighbor nodes of the matched users were extracted as new set of users to be matched.The identity matching was iterated repeatedly until all users were traversed.Experiments were performed on three truthful datasets and the results showed that the proposed method had better precision and recall rate compared with similar methods.(2)Social relationship type inferenceTo expand categories and improve accuracy of the inference result,a new method was proposed for inferring the social relationship types,which took mobile users' location data as research object and utilized the statistical data and semantic information of the spatio-temporal co-occurrence(STCO)areas comprehensively.The strength of STCO was calculated with the co-occurrence frequency of users,the entropy and discrete distance of STCOA.Based on the identification results of public place and user's personal location,the semantic information of STCOA was obtained.Then,the STCO vector was generated by combining the strength and semantic information of STCO.In order to get eigenvector corresponding to semantic feature,the semantic feature of STCOAs was extracted using the principle of"minimum relevance and maximum redundancy",according to category correlation and redundant information of the feature of STCO vector.Base on sample characteristics and requirements of social relationship type inference,the multi-class algorithm of support vector machine was selected and improved to construct classification model.According to the source of location data,two kinds of social relationship classification models of unisource location data and heterologous location data were built.Experiments was performed on truthful datasets and the results showed that:The unisource location data classification model could effectively infer four types of social relationship between mobile users such as family,colleagues,friends,and others.The accuracy rate was as high as 89.3%.Compared with the similar method,the precision and recall rate were increased by 7.9%and 6.6%respectively;Combined with user identity matching information,the heterologous location data classification model could get better user relationship inference accuracy than the unisource one.(3)Group discoveryThere is a problem that the social relationship and location attribute cannot be taken into account at the same time when using the existing methods to discover the groups of mobile users.A new method was proposed for overlapping group discovery based on inverse-label propagation algorithm(inverse-LPA)to solve the problem.Firstly,according to the location information of mobile users,the new method inferred the structure graph of social relationship before extracting the STCOAs.The STCOAs were used as position attribute labels to mark the graph.Secondly,the label graph was processed with the inverse-label propagation algorithm to remove companion-labels of nodes.With repeated iterations,each node preserved main-label of the groups when the state of the labels was stable.Finally,according to the user social relationship and node's main label under stable state,the groups of mobile users were divided and recognized.Furthermore,a multilevel inverse-LPA was given in combining with the type of user social relationships.Result of experiments on four truthful datasets showed that,the inverse-LPA took better account of social relationship and location attribute.It could detect overlapping group structure of mobile users more preferably than other methods.The multilevel inverse-LPA could improve the accuracy of the group discovery results with the information of user social relationship type.(4)Abnormal behavior detectionAiming at the problem of trajectory evolution and single-type of detection result in trajectory anomaly detection technology,an anomaly behavior detection method was proposed for mobile user based on location information,which comprehensively utilized the user's historical behavior pattern,group structure information,and behavior of close users.The method converted the location data into the STCOA,and further excavated the user behavior pattern and group structure information.On this basis,a multi-class anomaly detection model was constructed by random forest method,according to the five abnormal characteristics of historical behavior pattern anomaly,accompanying behavior pattern anomaly,STCOA discrete anomaly,STCOA aggregation anomaly and group attribute of abnormal users.This model could identify individual anomaly,group anomaly,location anomaly and event anomaly of mobile users.Experiments on real datasets showed that the proposed method could effectively identify the trajectory evolution behavior and detect various types of anomalies of mobile users.Compared with the similar methods,this method had higher recall rate and lower error rate.