An Efficient Framework for Crime Prediction Using Feature Engineering and Machine Learning

The growth of crime in society produces insecurity among people and severely impacts the country’s economic development. Understanding crime patterns is necessary to provide a proactive response to curb criminal activities. Most crime prediction models harnessing machine learning/deep learning techniques fail to exhibit significant results for the vague and inconsistent dataset. This is due to the non-consideration of suitable pre-processing and feature engineering techniques, resulting in an unreliable and inaccurate model. Hence in this work, an efficient framework for crime prediction using Feature engineering and Machine learning is proposed. Initially, the vague and inconsistent input dataset is pre-processed using Feature Scaling and Encoding input values. After pre-processing DBSCAN clustering is applied to extract geospatial features, which is used for improving model accuracy. Then SelectKbest library is used to select the 10 best features for the prediction model. Further, the SMOTE library is used to oversample crime types with the lowest sample size. Finally, for classifying the crime types, XGBoost and LGBM classifier models are evaluated. The proposed crime prediction model was validated by deploying an Amazon EC2 P3 instance in the cloud environment. The proposed model yields an improved accuracy of 62% for the XGBoost Classifier and 63% for the LGBM Classifier compared with KNN and Decision Tree.