| As the core technology of predictive policing,spatial-temporal(ST)prediction of crime has developed rapidly from around 2000 to the present.However,the existing work still has the problem that the data governance system is not clear and the model fusion ability and generalization ability are not strong.This paper has done the following work on ST prediction of crime:(1)We summarize the basic theory of ST prediction of crime at the beginning.We regard the ST prediction method of crime as a process combining the corresponding models to predict the ST distribution of the crime in the future,and deconstruct it into the relationship among three objects: case,ST backcloth and individual behavior.Then,from the point of view of the input factors of the prediction model,we sum up three currently mainstream methods,which are(1)the prediction method based on the information of cases’ ST location,(2)the prediction method based on the backcloth & the information of cases’ ST location,(3)the prediction method based on the individual behavior & the backcloth & the information of cases’ ST location,and summarize their mechanism in detail respectively.(2)Taking Beijing as the empirical research object,an environmental heterogeneous information network is constructed based on more than 3.5 million ST environmental data of six categories,such as POI/AOI,block form,business circle/life circle,street view images and road network.The data are divided into macro ST variables and micro cognitive variables according to the scale of road network structure.Artificial experience model and the deep learning-based representation learning model were used for feature extraction of different variables respectively.Finally,combining with the theft case data of Beijing in 2014 and 2017,feature fusion was carried out to obtain the native risk value of crime,and the modeling of ST background cognitive variables was completed.(3)Based on the ST location information of the case events,the location of the criminal ex-offenders and the associated case events,the heterogeneous information network of the social relations of the criminal ex-offenders is constructed.The artificial experience model and the deep learning-based representation learning model are respectively used to extract the features of the data.Finally,the ST risk perception variables in the ST prediction model of burglary crime are obtained based on the native risk value of crime based on the Journey-tocrime model,such as the travel probability of the criminal subject,the relationship characteristics of the subject,and the time-varying crime risk value.(4)Finally,according to the crime pattern theory’s ST hot spot transfer law,combined with CTR model of deep learning and the promotion method of integration learning,the ST prediction and correction of crime were carried out respectively.The best AUC value of burglary prediction could reach 79.72%,while the pickpocketing is 78.41%,and the best AUC value of burglary correction could reach81.14%,while the pickpocketing is 82.16%.Compared with the ST-GNN model which only considers ST location information of case events,it is improved greatly.Then,the interpretability of the input feature combination is analyzed,and the ST risk map is generated.Based on this,the prototype system of intelligent police patrol control is designed and implemented.At the same time,we present the model optimization needs to focus on improving the capacity to fuse heterogeneous data from multiple sources,and balancing the interpretability of models & predictive performance. |
