Research On Neural Network Weighted Regression Fusing Geographical And Attribute Space Features And Urban Crime

The analysis of spatial heterogeneity is of great significance for understanding the relationship between social phenomena and geographical elements.Geographically weighted regression is a local spatial statistical analysis technique,whose core is the measurement of spatial distance and the calculation of spatial weights.In geographically weighted regression models,Euclidean distance is usually used as the default metric.However,two points may be very close geographically,but their intrinsic properties are very different,so the Euclidean metric cannot fully express the spatial proximity relationship.In terms of spatial weight calculation,due to the mathematical structure of the spatial weight kernel function and the limitations of the model optimization method,the spatial weight calculation is inaccurate,which limits the spatial analysis and modeling expression capabilities of the geographic weighted regression model.In order to solve the above problems,this paper focuses on the differences between the intrinsic attributes of geographic elements,aiming to construct a new spatial relationship measurement method that integrates spatial features and attribute features to fully express the spatial proximity relationship.Specifically,this paper uses the attention mechanism to extract the spatial features and attribute features of the fitting points,deeply explores the internal relationship between spatial features and attribute features,and realizes the space-attribute distance measurement;utilizes the powerful learning ability of deep neural networks to construct a geographically spatial attribute neural network weighted regression,achieving accurate calculation of spatial weights;takes the modeling of the spatial non-stationary relationship of violent crime and property crime in Chicago,USA as an example,and validate the geographically spatial attribute neural network weighted regression method.The main content of the paper is summarized as follows:(1)With the goal of "extracting spatial and attribute features" and "fusing spatial and attribute features to form effective spatial relationship measurement methods",a spatial attention module and attribute attention module were constructed,and a spatial attribute weighted neural network was designed to form a systematic framework of a geographically spatial attribute neural network weighted regression model.Statistical inference methods were also provided,achieving innovation in modeling spatial non-stationary relationships.(2)The training framework of geographically spatial attribute neural network weighted regression model was built;Thoroughly studied the relationship between spatial features and attribute features,and designed four spatial attribute weighted neural networks based on different feature extraction and fusion ideas;The model optimization strategy were designed to improve the model’s ability to detect spatial heterogeneity.(3)Taking the spatial non-stationary modeling of violent crimes and property crimes in Chicago,USA as an example,a comparative experiment was designed between a geographically spatial attribute neural network weighted regression model and multiple geographically weighted regression models to verify the high performance and reliability of the geographically spatial attribute neural network weighted regression model;The commonalities and differences in the driving forces of violent crimes and property crimes have been studied,revealing to some extent the causes and spatial changes of crime.