Prediction Method Of Theft Crimes In Urban: An Integrated Model Of LSTM And ST-GCN

Larceny is a common crime against property that generally has a large number and needed a lot of police resources.The accurate prediction of the temporal and spatial distribution of larceny plays a supporting role in improving the efficiency of police work such as control,investigation and patrol.At present,researches on crime prediction methods at home and abroad are mostly limited to large-scale of temporal and spatial,which have limited accuracy and precision,and the forecast results are generally insufficient in guiding the actual police work.In response to those issues,this paper carries out study on urban theft cases.By analyzing and extracting the sequence and spatial-temporal characteristics of theft cases and combine them with environmental factors,in unit of day and community,an integrative model:LSTM and ST-GCN is established for urban theft crime prediction.The contents and achievements of this paper are listed as follows.(1)Established an integrative model:LSTM and ST-GCN for urban theft crime prediction.the model can be categorized into three modules—spatial-temporal feature extraction module,temporal feature extraction module,and feature integration module.First,in order to detect crimes in each community,the temporal feature extraction module is built based on the LSTM network.Then,the spatial-temporal feature extraction module is a combination of GCN and ST-Res Net(ST-GCN)to extract the transition of crimes in space over time.Finally,the feature integration module employs GBDT model to integrate the predicted values from the spatial-temporal feature extraction module and the temporal feature extraction module.Taking"day"as the temporal scale,the model extracts the spatial-temporal characteristics such as proximity,periodicity and trend.Taking"community"as the spatial scale,the model puts forward the prediction method based on spatial topology structure,and takes variables such as weather,"holiday"and"weekend"as external characteristics to predict the spatial-temporal distribution of urban theft crime.(2)verified the prediction model of urban theft crime.Based on the statistics of all theft data in all 77 communities of Chicago from 2015 to early 2020,this paper takes the data from2015 to 2019 as the training set to train the above model.In the test set(theft data from January1 to March 10,2020),the RMSE and R~2of the fusion model proposed in this paper are 1.03 and0.84,which are better than ridge model(RMSE is 4.04 and R~2is 0.48),random forest regression model(RMSE is 3.91 and R~2is 0.53)and LSTM model(RMSE is 1.59 and R~2is 0.75).In addition,this paper also puts forward a prediction model based on the combination of LSTM and GCN to classify cases degree,thus the reliability of the prediction results can be verified in application.The classification and prediction accuracy of this model for each community is more than 80%,and the highest can reach to 100%.The prediction model of urban theft based on LSTM and ST-GCN proposed in this paper is expected to provide support for policing decision-making under various business scenarios,such as public security risk prevention and control,police resource allocation,community policing work,large-scale event security and so on.