Crime Data Analysis Using Machine Learning Models

Abstract

Crime statistics in Ecuador show us that in recent years the number of cases for different types of crimes has increased. Although the different state entities have criminal data, analyzes are not always carried out to predict new cases. This work proposes an analysis of the information based on automatic learning algorithms that allows extracting knowledge about the relationships between the different variables that affect criminal acts. These results can be used as tools for the country's authorities and organizations to better control and prevent crime. Using machine learning algorithms, crime counts by province can be predicted using techniques that are based on multiple regression or other techniques. Using monthly counts of different types of crimes over several years, three machine learning algorithms are implemented: Multiple Linear Regression (MLR), Decision Tree Regression (DTR), and Random Forest Regression (RFR). These models are trained and tested for use in predicting new crimes, especially rapes, burglaries, and personal thefts. The R-squared, adjusted R-squared, and root mean square error (RMSE) metrics are used to evaluate and compare the proposed regression models. The results show that the RFR model achieves a better fit to the data with an adjusted R-squared value of 0.965746 for the case of home burglaries and a value of 0.974088 for thefts. In addition, this model presents the lowest RMSE value for the three types of crimes. The best adjusted R-squared value for the rape case was obtained using the MLR model with a value of 0.929960. The most affected provinces in absolute counts are Guayas and Pichincha, whose crime levels remain at alarming levels.