Mobile Users’s Spatial-temporal Trajectory Prediction

With the rapid economic and social development and the arrival of the era of big data,features such as social informationization,dynamic security,crime intelligence,politicization of rights protection,normalization of anti-terrorism and other characteristics have become increasingly prominent.The traditional public security policing model is difficult to adapt to new forms and new Task requirements.Faced with the complicated and changing domestic and international social environment and challenges in the era of big data,we need to firmly establish big data thinking,innovate big data technology,and deepen the application of big data.Accelerating the transformation and upgrading of public security mechanisms is a major task and responsibility of the public security organs.To this end,according to the unified deployment of the municipal bureau,the Public Security Bureau accelerates the implementation of special projects for the application and construction of public security big data,and strives to create a modern policing mechanism for public security under the era of big data.This research originated from the demand of the company where the article was intended to undertake public security projects in Shanghai.The Ministry of Public Security proposed the establishment of a rogue criminal path prediction model.Due to the relatively slow progress of coordination of crimes and other data,the general users of telecommunications were first used as the research objects and their user space-time was studied.The trajectory data proposes a feasible general algorithm framework,which is then used in specific contexts,embodying ideas from deduction to induction.In view of this situation,this paper uses the mobile phone signalling data captured by the operator’s base station and combines external data such as user information,weather data,Chinese holidays and commercial tag information as the basis for data analysis,and adds Baidu maps and the Gaode map API to find each one.The latitude and longitude corresponding to the base station integrate multi-source heterogeneous data sources to better reflect the entire space-time trajectory of mobile users,IV refine the strong feature domain,and lay a good foundation for later data modeling work.The innovative work of this paper on this topic consists of the following three types:(A)positioning algorithm optimizationVarious communication scenarios include 2G,3G,and 4G.In order to cover these scenarios,base station positioning is performed to creatively integrate these positioning data OIDD,PCMD,and LTE.And based on the delay-based triangulation algorithm TDOA,multiple base stations receive signals at the same time,and the distance from each base station is calculated according to the signal arrival time so that the error does not exceed 50 meters.(B)Improve Training ModelSince the location data of Shanghai Telecom subscriber base stations is very large,how to effectively carry out model training is the primary consideration.In this paper,the geospatial space is meshed,and the relaxation variable is added to solve the sparseness problem of the boundary data.Both the model and the hyperparameter selection can be run in the grid.Through this subset,the performance of the model can be known in a short time.In the feature construction process,counting features,sorting features,feature combinations are added,and time features are captured from 3 cycles on the basis of time discretization,which effectively reduces the model error.(III)Model fusionEach single model will predict a high probability of ten landmarks per user.This article uses the default dictionary class to combine the probabilities of different models and uses the first three higher probability landmarks as the forecast output.This article first validates the set integration to see if the model portfolio has better performance.Then,using this combination to run a fusion over the entire data set,test fusion conditions at the same time whether the low correlation is satisfied,and then use this combination to weight the Bayesian estimates to get higher model accuracy.The final experimental results show that the fusion model has higher accuracy than the single model,the MAP3 score of a single model is about 0.51,and the Bayesian optimizationmodel of the fusion model will reach about 0.82.After several experiments,the feature engineering and the combined feature training model were continuously improved.Hyper-parameter optimization made this result stable.The key to model fusion lies in the model difference.Differences are mainly reflected in different models,different parameters and training of the same model.The data is different.Before each model fusion,it is necessary to refer to the correlation of each model.You can use Cosinesimilarity or Person coefficient evaluation,or you can use the MIC maximum information number to test,so as to effectively avoid overfitting.Of course,there is still much room for improvement in terms of feature work and adjustment.Judging from the current results,this model shows a high degree of feasibility and effectiveness,which will help the public security agencies to investigate the security situation,find out the security focus,deploy the police force scientifically,and evaluate the performance of the work,so as to achieve the ultimate goal of aid decision-making.