基于集成特征选择的盗窃案件预测方法

盗窃类案件是公安机关较为棘手的一类犯罪，呈现高发低破态势.提前预测发案情况是预防该类型犯罪的有效途径，因此对预测盗窃犯罪提出了一种以Bagging方法为基础、基于特征选择准确度和差异性双重考量的集成学习算法，根据集成学习器好而不同的原则，构造由异质基学习器集成的特征选择器，实现对影响盗窃犯罪发生因子的有效选择，使用更少维度的特征数据集提升犯罪预测的效率和准确度.实验结果表明，提出的SEFV_Bagging算法具有较好的泛化能力和稳定性，在测试数据上表现出的预测准确度也较为理想，且算法无需根据先验知识设置所选特征子集维数，在盗窃犯罪数据分析预测领域应用中有较为明显优势. 

Theft Prediction Method Based on Ensemble Features Selection

Theft crime is a difficult problem which shows a high occurrence and low breaking situation. It is an effective way to prevent the crime by predicting the cases in advance. So a new method was proposed based on bagging, following standards of accuracy and differences in feature selections, with the principle of high accuracy rate and difference rate. Heterogeneous learners were used to construct an ensemble learner to identify the occurrence factors, then the efficiency crime prediction was improved with less dimensions of factors. The results show that the proposed SEFV_Bagging algorithm can provide better generalization ability and stability, also its prediction accuracy is better. In addition, the algorithm needn''t transcendental knowledge to set the feature subset dimensions manually, which shows obvious advantages in the application of criminal data analysis and forecasting.