User Re-identification Based On Mobile Behavior

Pedestrian re-identification refers to a new technology that uses computer vision,pattern recognition,and other technologies to extract the feature of pedestrian images across planes,so as to carry out personnel matching.At present,pedestrian reidentification is mostly based on video and image data.However,with the rapid development of information technology and the popularity of mobile devices,the use of trajectory data generated by mobile users for re-identification also has research value,which lays a technical foundation for user travel route selection,urban traffic planning,and other scenarios.In this paper,call details records and social network check-in data are used as data sources,and the technical advantages of deep learning in semantic feature extraction and analysis are mainly studied.The statistical analysis of different trajectory data is carried out.Combined with the characteristics of the data,the time series contained in the trajectory and the geographical coordinates in the position of interest points are modeled.Compared with the traditional method,the proposed method has higher recognition accuracy and can fully excavate the potential correlation between different interest points in sequence data.In addition,in order to further improve the performance of the user re-identification method,this paper innovatively uses the retrieval-based method to realize the query and discrimination of users,which overcomes the problems of poor recognition accuracy and low efficiency in the past classification method when the number of users is large.The experimental results have good performance in multiple datasets.(1)In order to better understand the behavior regularity of mobile users,this paper analyzes and counts the above datasets by using the data mining preprocessing method based on the call detail record data in Shanghai and the open-source data in social networks platform.The study shows that most users' daily travel distances in the call details record data with dense sampling are between 10 km and 50 km on average.By using the visualization of the thermal map,it is concluded that the user event distribution follows the characteristics of the time cycle.In addition,the user's travel range gathers in the form of clusters in the geographical space,and it is difficult to find clear rules in the spatial and temporal dimensions of the sparsely sampled check-in dataset.(2)In the trajectory similarity mining task,most traditional methods calculate the trajectory similarity through statistical strategies such as frequency and distance.These methods are simple but ignore the previous and later relationships in the trajectory sequence.In this paper,we use the deep learning model to extract the characteristics of the trajectory based on the analysis of the spatial and temporal distribution of the users for the call detailed record data with dense sampling.Firstly,the trajectory is transformed with the graph structure,and the node2 vec algorithm based on the graph structure is used to pre-train the embedded representation of each base station site,so as to obtain the input sequence of the subsequent model.Then,this paper uses the hierarchical attention network to learn the behavior patterns of users at different times of each day,and constructs a multi-connection structure from position to sub-trajectory and then to the whole trajectory.In order to make full use of the spatial information in the sequence,the convolution network is further used to extract the geographic information characteristics,and the two characteristics are fused and the user identification is completed through the multi-classifier.The experiment shows that the recognition accuracy of the model in this paper reaches 95 % and 90.25 % under 400 users and 800 users,respectively,which is significantly improved compared with the traditional similarity measurement method.(3)When the classification method is used to re-identify the trajectory of largescale users when new users join the model that needs to be retrained and a large number of tagged user trajectory information is required,it is usually difficult to obtain this information.In this paper,based on the check-in data of social networks,a network model based on the Siamese-Transformer is proposed based on the idea of retrieval.The model is divided into two modules: the discriminant module and the retrieval module.In the discriminant module,the encoder in Transformer is used to learn the global trajectory information.Compared with the commonly used recurrent neural network,Transformer reduces the performance degradation due to long-term dependence by overall processing trajectory information.Then,a joint attention mechanism is introduced to capture the potential correlation between different trajectories.Thus a more comprehensive and high-level understanding of the trajectory is generated.The retrieval module selects the candidate user set corresponding to the re-identification trajectory based on the relevant information between the user and the location and combines the discriminant module to obtain the attribution user of the anonymous trajectory based on the KNN algorithm,so as to complete the user reidentification task.In the experimental analysis part,the performance indicators of different datasets,different user numbers,and different re-identification methods are compared.The results show that the proposed model is better than the existing methods.In addition,in order to verify the effectiveness of the joint attention module,the ablation experiment proves that the module can further improve the accuracy of user reidentification.