Enhanced Approach for Crime Predictive Analysis Using NCRB dataset

Abstract

Crime prediction is a significant aspect of contemporary policing as it allows allocating resources proactively, planning, and making decisions based on evidence. The statistics of crime as released by National Crime Records Bureau (NCRB) in India have been mostly employed in descriptive analysis and not used in prediction. This paper suggests a machine learning-oriented model to predict crime patterns with the help of NCRB data both national and state. To create a homogeneous analytical dataset, data regarding crime on a state by-state basis (2011- 2022) and national, subcategory-level data (2000- 2022) were assembled and reconciled using annual NCRB reports. As the base learners, two gradient boosting models (XGBoost and CatBoost) were used, and their prediction was stacked by a stacking ensemble model with Ridge Regression as the meta-learner. More complex methods of feature engineering such as lag variables, trend indicators, differential measures and spatial correlation features were added to ensure continuity in time in crime patterns and relationship between regions. Experimental analysis shows that the ensemble model has better predictive accuracy than the individual models with lower values of error and higher value of coefficient of determination. The trained model was also used to produce long-term crime predictions over the next three years (2023-2030), as well as statistically significant crime hotspots based on z-score analysis. The visualization dashboard was made interactive to show the past trends and projections as well as the spatial hotspots distributions to enhance understandability. It was shown that ensemble learning in conjunction with structured feature engineering is a promising method to forecast crime on large-scale and aid the development of data centered policy and policing decisions in the Indian case.