基于犯罪空间分异和建成环境的公共场所侵财犯罪热点预测

机器学习是当前犯罪热点预测的主流方法,随机森林算法因需要的数据量较小、有较好的预测能力和预测精确度、且有较高的可理解度,更是被广泛应用,代表地理环境和建成环境的多源数据也被广泛用于模型改进的尝试实践中,但这些实践都只考虑研究区整体的预测精度变化情况,并未区分不同区域犯罪热点预测结果的差异及其原因。因此,本文以公共场所侵财犯罪为例,根据历史犯罪分布情况及过往犯罪热点分布规律,将研究区分为稳定高发热点网格、较高发热点网格、偶发热点网格及非热点网格这4类,并依据社会失序理论、日常活动理论和犯罪模式理论,选取城中村范围、路网密度及POI（餐饮、娱乐、商场3类设施）密度这3个具有代表性的协变量加入到随机森林预测模型中,探讨预测结果精度的变化情况。根据2017年26个双周的犯罪热点预测实验的预测结果,得到以下结论：加入协变量后,研究区整体、稳定高发热点网格及较高发热点网格的预测精度都有不同程度的提高,分区模型的精度显著高于整体模型的精度,说明考虑空间分异对提高模型精度起重要作用。

Hotspot Prediction of Public Property Crime based on Spatial Differentiation of Crime and Built Environment

Machine learning is the mainstream method for crime hotspot prediction. As a popular machine learning algorithm, the random forest algorithm is widely used in the construction of crime hotspot prediction models because of its ability of handling sparse data, and reliable predictive capability and accuracy. A number of studies use multi-source data representing the geographical environment and built environment to train and construct crime hotspot prediction models. Some are theory-driven, while others more data-driven. Most crime prediction models are global models, by constructing a single model for the entire study area. These models do not fully consider the spatial variations of crime and the built environment, as well as the varying relationship between crimes and the built environment. This paper aims to fill in this gap, using public property crime as an example to demonstrate that crime prediction models can be improved by incorporating the aforementioned spatial variations and spatially varying relationship. Firstly, according to the distribution of historical crime events and the distribution of past crime hotspots, the research area was divided into four subareas: stable high-heat grids, high-heat grids, even-hot grids, and non-hot grids. Then, according to the social disorganization theory, routine activity theory, and crime pattern theory, the three covariates including the urban village, the road network, and POI (catering, entertainment and shopping malls as crime attractors and generators) were used as the covariates representing the surrounding built environment. The random forest prediction model also used historical crime data for training and validation. Different models were created for the whole study area and each of the four subareas. The results of 26 bi-week crime hotspot prediction experiments in 2017 were compared, showing that, after adding the three covariates representing the built environment, the prediction accuracy of the entire study area, stable high-heat grids, and high-heat grids were all improved. More importantly, the subarea models were substantially more accurate than the whole model. These findings strongly endorse that incorporating spatial differentiation of crime and the built environment plays a critical role in improving the performance of the prediction models. The majority of the hotspots coincide with commercial facilities that serve as crime generators or attractors. Thus, crime prevention and control should target urban villages and the areas where road densities are high. Further, the differences in subarea based models also suggest any crime fighting strategies should be adjusted to fit each local subarea, to achieve the greatest efficiency.