A Crime Pattern Detection using Predictive Policing with Random Forest based on Synthetic Minority Oversampling

In recent days, major cities have created several problems in terms of public safety and early detection of crime which is an important part to reduce the impact. In existing Crime Pattern Detection (CPD) models such as impact learning and Random Forest (RF) majorly struggles with high dimensional data. In this paper, predictive policing with Random Forest based on Synthetic Minority Oversampling (PP-RF-SMOTE) is proposed for effective CPD with reduced class imbalance by producing artificial marginal samples in order to overcome sample minority class. At first, the data from Chicago open data portal is preprocessed with one-hot encoding to convert the categorical values into binary vector representations. Then, the min-max normalization is utilized to fit the binary vector representations into a certain range. Then, RF is introduced to select the optimal features from normalized binary representations. The SMOTE technique generates synthetic minority samples to oversample them in a minority class to reduce the class imbalance in RF feature selection. At last, XGBoosting classifier is utilized to classify the crime patterns into binary labels as crime or non-crime. From the results, the proposed PP-RF-SMOTE model showed improved results in contrast to existing impact learning by providing accuracy of 90.35%, precision of 92.17%, recall of 93.78% and F1-score of 92.96% respectively.