Research On The Spatio-Temporal Prediction Method For Urban Crime Risk

Spatio-temporal crime risk prediction plays an important role in societal security risk prevention and management.However,the interpretability and accuracy of traditional patio-temporal prediction model are relatively poor.In order to improve these issues,with crime risk represented by crime incident numbers,empirical studies were conducted on the crime records collected in a large city in the north of China,to investigate the spatio-temporal patterns of crime risk,then a time series prediction method based on FFT and Kalman filter as well as a spatio-temporal prediction model integrating FFT and LSTM-GCN were proposed,respectively.To verify the performances of the methods proposed in this study,case studies were carried out on three kinds of datasets,including the dataset of all types of crime incidents in CY District of a large city in northern China,the dataset of assault incidents in Chicago and the dataset of theft incidents in Chicago.The main conclusions of this study are as follows.（1）Using five kinds of crime rates（crime incident numbers per million people）and five types of climate variation records ranging from 2005 to 2016,the temporal patterns of crime risk in a large city in northern China were explored,then the impacts of climate variations on crime rates were investigated.Meanwhile,using 8 kinds of social-environmental factors and the crime incident numbers collected in 2019 and 2020,the spatial patterns of crime incidents in the CY District of the city were examined.The results show that,（1）in terms of the seasonality of crime rates,minimal violent robbery（MVR）,robbery,assault and rape have seasonality.Strong positive relations are observed for temperature-to-MVR,temperature-to-assault,and temperature-to-rape,with R~2 higher than 0.75.（2）In terms of the daily variations of crime rates,temperature,rainfall and haze are key factors.The relations are all positive for temperature-to-assault,temperature-to-rape,temperature-to-homicide,rainfall-to-robbery and rainfall-to-MVR.（3）Local spatial autocorrelations are observed for the crime incidents in CY District,and the spatial distributions of crime incidents are significantly correlated with GDP and other four social-environmental factors,with R~2 higher than 0.1.（2）A time series prediction method based on FFT and Kalman filter was proposed in this study.By extracting the long-term trend and high-energy periodic components from the time series of crime incident numbers,and predicting those components with machine learning models,this method can provide a predicting interval for crime incident numbers,which is composed of predicted baseline and tolerance interval.To verify the performance of this method,case studies were conducted on three datasets which are referred above.The results show that,（1）the long-term trend and periodic components can be well predicted,with R~2 all higher than0.85.（2）The PICP index of interval prediction results on the three datasets are all higher than96%.The interpretability and practicability of this method are higher than traditional time series prediction method.（3）A spatio-temporal prediction model integrating FFT and LSTM-GCN was established in this study.The FFT module can provide parameter basis for LSTM-GCN module through spectrum analysis,and LSTM-GCN module can predict the spatio-temporal distribution of the crime incident numbers.To verify the performance of the model,case studies were carried out on three datasets referred above.The results show that,the MAE of the prediction results on the three datasets are 1.01,1.21 and 2.37 respectively,and SMAPE ranging from 30%to 50%,which is lower than LSTM-STGCN,STGCN and other models.The model can perform better in crime risk prediction tasks with sparse data than other spatio-temporal prediction models,because of its higher accuracy,stronger robustness and lighter structure.（4）Specific statistical patterns can be observed from the crime incident numbers in a certain spatio-temporal range,and these patterns will be more significant as the spatio-temporal range expands,which provides theoretical basics for crime risk（crime incidents number）prediction tasks,but there is a theoretical upper limit for the accuracy of the task.