A Spatio-temporal Prediction Model Of Crime Based On Knox And Grid Management

Since China’s reform and opening up,the government’s emphasis on regional development strategy,speeding up the process of urbanization and population migration have led to the complexity of the crime situation.All kinds of domestic crimes(such as robbery,theft,conflict,hit and run,etc.)have shown a trend of "high incidence and low breakthrough",bringing unnecessary losses to the country in human,financial and material resources It has seriously affected the normal order of the society.As a common social phenomenon,the distribution of frequent crime in time and space is not random,but shows historical regularity in its development process."Daily activity theory" points out that criminal activities are always closely related to the surrounding physical environment.Moreover,with the deepening of economic reform,the public security department has accumulated a large number of crime data,which can be used to support the research and application of crime prediction.So far,scholars at home and abroad have proposed a variety of crime prediction methods based on the characteristics of crime spatial-temporal distribution,which can better solve the problem that traditional artificial crime analysis and prediction methods are difficult to systematically analyze real-time data.However,these crime prediction methods only calculate the interval length of discrete case points in different dimensions for statistical analysis,and to some extent ignore the characteristic property of near repetition of crime.At the same time,researchers usually ignore the highly uneven distribution of crime data in the global geographical region,which leads to a large proportion of "zero crime" in the input features,and it is difficult to identify and extract the crime features,which makes the prediction results of the model sparse and serious.On the other hand,in the initial stage of model training,the traditional deep learning algorithm often uses prior knowledge to determine its initial parameters,which directly affects the construction of network structure and makes the model prediction results unstable.In order to solve the above problems,this paper takes four types of frequent cases(theft,assault,assault and criminal damage)in Chicago as the research object,and uses the spatiotemporal aggregation calculation method in infectious disease research for reference,proposes a spatio-temporal crime data grid prediction model based on the combination of improved Knox algorithm and improved QPSO optimized DBN network.Firstly,the Knox algorithm based on the average nearest neighbor distance is used to avoid the problem of insufficient autocorrelation analysis between cases;secondly,the grid geographic information management method and the K-means low-density spatio-temporal cluster under sampling method based on the centroid are used to alleviate the imbalance of the existing crime data distribution;finally,the improved QPSO algorithm based on the proportion of the length change rate of the particle well is used to solve the problem It improves the stability of traditional DBN network model.This paper scientifically selects the initial parameters of the crime spatio-temporal prediction model,and adds geographical features in the input of the model,which improves the utilization rate of the crime related data set of the model,and realizes the analysis of the spatio-temporal distribution of four types of cases and the phenomenon of near repetition of crimes.On the basis of the above results,this paper visualizes the prediction results,in order to provide reference for the accurate command of police decision-making and the deployment of crime prevention and control.The research contents and achievements of this paper are as follows:(1)This paper innovatively applies Knox algorithm of spatio-temporal aggregation degree of infectious disease research to the field of crime,studies the near repeatability scale of crime space-time,and proposes a Knox algorithm based on the average nearest neighbor distance,which is named Mnd-Knox algorithm.On the one hand,it innovatively adopts the Knox algorithm,which is different from directly using the original number of crimes On the other hand,it can better fit the development law of frequent crime cases.Considering the closer relationship between adjacent cases and distant cases,it has better effect in predicting frequent cases.(2)This paper introduces the idea of urban grid management.On the premise that the internal autocorrelation between crime points is completely preserved,the pretreated crime data points are put into the three-dimensional spatio-temporal interactive grid structure.Combined with the kernel density estimation method,the spatio-temporal data information in the unit grid is counted,and the K-means low-density spatio-temporal cluster undersampling method based on centroid is used to balance the majority of samples from the technical level and the number of a small number of samples to solve the problem of uneven distribution of crime data.(3)In this paper,by integrating the characteristics of crime related environment,determining the near repeatability threshold of crime spatio-temporal based on Mnd-Knox algorithm,constructing three-dimensional spatio-temporal interactive grid,K-means lowdensity spatio-temporal cluster undersampling method based on centroid and DBN-DNN network design,we establish the prediction model of crime spatio-temporal distribution based on DBN-DNN network,and effectively learn and train the cross spatio-temporal domain crime cases In addition,the performance of the prediction model and its prediction results are comprehensively evaluated and analyzed.(4)This paper proposes an improved QPSO algorithm based on the proportion of the change rate of the length of the particle potential well,and uses the algorithm to optimize the DBN parameters to establish a crime spatio-temporal prediction model,which is named the crime spatio-temporal distribution prediction model based on L-QPSO-DBN network.The model further optimizes the model in(3),avoids the random problem of determining the network structure parameters based on prior knowledge,and improves the efficiency The accuracy and stationarity of the standard DBN-DNN model are analyzed.(5)Based on the prediction results of L-QPSO-DBN network,this paper visualizes the hot spot map of crime prediction results by using plotly express in Python.The hot spot map divides and counts the number of crimes by grid,and divides the prevention and control level according to the predicted number of crimes,which provides the basis for the deployment of police daily patrol work.The experimental results of four types of frequent cases in Chicago show that the spatialtemporal crime data grid prediction model based on Mnd-Knox algorithm and L-QPSO-DBN network can effectively realize the crime prediction of the study area.The average absolute error of the four types of cases prediction model obtained by data feature optimization reduces by 88.56%,and the DBN improved by L-QPSO algorithm is used.The standard deviation of prediction accuracy of the network model is reduced by 45.17%,and the overall performance of the model is good,which can achieve the expected goal,especially when it is used to analyze the types of cases with significant autocorrelation.