Research On Crime Prediction Technology Based On Deep Spatio-temporal Networks

Spatio-temporal data mining is one of the cutting-edge topics in the field of deep learning.Crime prediction is an important branch of it.However,crime prediction faces some problems.Spatial research is monolithic and lacks the capture of long-term dependencies.The accuracy of hour-level predictions of fine-grained crime is low due to the high sparsity of data.To address the above issues,this paper focuses on crimes such as theft and constructs deep spatio-temporal networks to predict future crimes.A crime prediction model based on improved LSTNet is constructed for coarse-grained crime data with a temporal granularity of months and a spatial granularity of communities or above.First,the model adds an STL module,which is used to decompose the trend component T,the seasonal component S and the residual R.Then,the three networks in LSTNet,LSTM,SkipRNN,AR is adjusted to a parallel structure,so that the three components T,S,and R can be fed into each of the networks.The output of the three parallel networks is input to the convolutional layer.Finally,spatial attention and category attention mechanisms are introduced to capture the correlation between other geographic units and the target crime,as well as the correlation between other categories of crimes and the target crime.The model addresses the short of capture of long-term dependencies and spatial singularity.Experiments on crime data in Chicago,USA,showed a mean RMSE of 2.968,which outperformed other models.For fine-grained crime data with hourly time granularity and grid spatial granularity,a crime prediction model based on improved ST-ResNet is built.First,the crime data at each moment is transformed into a pixel distribution of images.Then,a temporal feature extraction module with adjustable depth is designed.It extracts the temporal feature of crime data by layer overlay method.And the above time features are fed into four parallel residual networks to capture the spatial correlations of near,middle and far distances.Finally,the outputs of the four residual networks are fused,and correlation factors such as weather and holidays are added to the model.The model addresses the problem of poor model performance due to the high sparsity of finegrained crime data.Experiments on Chicago crime data showed a mean hit rate of 58.51% of crimes when the predicted area was 10% and a mean hit rate of 78.34% when the predicted area was 20%.Both the mean hit rate and PAI outperformed the other models.It was applied and validated in T-cities in China,which achieved good results and proved the feasibility of the model.