A Study Of Multi-scale Spatiotemporal Prediction Method That Takes Sparsity Into Account

The acceleration of urbanization has made the public security problem exceptionally prominent.In response to public security problems,it is not only the task of public security departments to implement prevention and control,and maintain the safety of human life and property,but also the top priority of scientific research.The enhancement of computer hardware and software equipment,city-related data are acquired in real time,and the development of machine learning technology,especially deep learning technology,gives the possibility of crime prediction.Crime hotspot detection technology and prediction technology have continuously gained progress.It is of great significance and research value in developing police strategies,inspection routes,timely prevention and control,and reducing urban crime.To explore the complex spatiotemporal hotspot dependence patterns of urban non-uniform data,more and more spatiotemporal hotspot prediction models have been developed.For example,early hotspot mapping focused on identifying and discovering areas with high concentrations of events,specifically how to spatially interpret the location,size,or other aspects of clusters.Among them,location-oriented and type-oriented crime hotspot mapping has gained more attention in business and academia because it provides protective analysis and policy development for public safety,but not limited to it;hotspot mapping is also a fundamental form of crime prediction.Prediction of crime hotspot mapping has shown to be of increasing importance.Existing prediction models based on rich data features or studies at large spatiotemporal granularity are still inadequate for prediction of sparse types of crime hotspots.On the one hand,crime data exhibit non-uniform,random and sparse spatiotemporal distributions at the city scale.Very little modeling is considered from a data perspective,i.e.,the bias introduced by the large number of zero values.On the other hand,single-scale models are unable to capture nonlinear crime features.There are neglected spatiotemporal multi-scale of the model,which poses a challenge to spatiotemporal pattern prediction modeling.In this paper,the public dataset of sparse Chicago assault crimes is used as the data source that divides multi-level spatial grid units for prediction experiments,and research is as follows.(1)Research on gridding and rasterization for multi-level spatiotemporal point data.The study of inner-city space requires more detailed and finer granularity.Hence,we adopt a multi-level partitioning method while satisfying the hierarchical principle of pyramid to divide the regular spatial grid units with different levels of coarseness and fineness.In addition,the number of events is calculated unit grid and unit time based on the programming idea.By generating a three-dimensional tensor dataset,it is used as an input to the prediction model.(2)Research on spatiotemporal feature mining of point events.Mining the spatiotemporal feature of event data can both help understand the hotspot distribution pattern and guide prediction studies.The first for sparsity,the sparsity factors and the geographic distribution of spatiotemporal crime sequence density are explored.The next for temporal and spatial regularity patterns,different analysis methods are used to explore spatiotemporal distribution features and patterns of the data by targeting the temporal trends,periodicity,spatial autocorrelation,aggregation,etc.,so that the differences in their distribution can be understood.And then,we explored the spatiotemporal hotspot patterns of crime through kernel density,spatiotemporal cube,and emerging hotspot analysis to guide the research of crime prediction problems and the construction of spatiotemporal prediction model.(3)Research on multi-scale neural network spatiotemporal prediction models.Existing research methods such as deep learning and graph neural networks are widely used in continuous data,but in discrete data,there are problems such as "fixed receptive field","zero value inflation","more complex spatiotemporal dependence".Therefore,this paper proposes a multi-scale spatiotemporal hotspot prediction model ST-HGNet for the model with fixed receptive field and zero-value inflation caused by spatial sparsity,which mainly explores the rationality and accuracy of the model from three perspectives.For spatial dependence,a hierarchical gated convolutional block is designed to enhance the sensitivity of spatial features through multi-scale representation,and the perceptibility of sparse tensor through gating.For temporal dependence,setting up the temporal closeness component and the periodic component for temporal trend mining.For spatiotemporal feature fusion,using parameter matrix fusion to dynamically capture complex spatiotemporal dependencies and predict hotspots.(4)Example analysis of crime prediction.The Chicago Assault Crime Type is selected as the case study,and several advanced deep learning models are selected as the baseline models for experimentation and evaluation.With the help of common evaluation metrics,such as RMSE,mean hit rate and PAI are used to evaluate the prediction results of different models.Meanwhile,ablation experiments are also set up to verify the ability of different component connection methods to capture spatiotemporal features.The contrastive experiments are used to find the most suitable prediction model for the case data.Finally,the hotspot prediction results are displayed with visualization analysis,which makes the experimental results more intuitive.