運用資料探勘方法預測 城市之犯罪類型

Abstract

犯罪行為可以透過科學研究方法發現犯罪特徵，並據以描述、解釋、建立區域犯罪預測模型，進而將犯罪加以抑制。犯罪預防策略上，大多依賴傳統巡邏臨檢、提高見警率等方式以預防犯罪發生，欠缺科學性評估，無法確切顯示犯罪之空間與時間分佈，無法有效提升犯罪偵防之效益。 本研究運用資料探勘技術，針對美國舊金山城市警察局的報案紀錄，由時間、空間等屬性資訊來預測發生的刑案類別。首先透過資料視覺化的方法，呈現犯罪類別與屬性之關係，找出影響案件發生的重要屬性，再分別以決策樹、隨機森林、貝氏分類演算法建立預測模式、預測發生的刑事案件類別，實驗結果顯示隨機森林的表現較佳。

Crime behavior can be analyzed through scientific methods to discover crime characteristics, which can be further used to describe, explain, and build regional crime prediction models to deter crimes. In the prevention of crimes, police departments strongly rely on traditional patrol and increasing police visibility. However, these enforcements can neither be scientifically evaluated, nor can present the crime distributions in space and time. Accordingly, it is not effective in preventing crimes. This research applies data mining techniques to analyze the crime reports of San Francisco and build prediction models to predict the category of crimes based on the crime information in space and time. Data visualization methods are used to display the relationships of crime categories and attributes, and identify the important attributes that affect the occurrences of crime cases. Decision tree, Random Forest and Naive Bayes classification methods are adopted to build the prediction models for predicting the category of criminal cases. The experiment result shows that the performance of Random Forest is the best among the classification models for predicting the crime category.