A Study on the Prediction of Crime in Chicago Based on a Two-Stage Machine Learning Approach

By forecasting the occurrence and frequency of crimes within a specific area, authorities and the community can prevent injuries and property losses. The study utilizes crime data from Chicago, USA, covering the years 2020 to 2024, classifying incidents into six categories: robbery, theft, burglary, assault, criminal damage and sexual violence. Each crime type is represented as either 1 or 0, signifying the presence or absence of a predicted crime. The data is then analyzed using CATBOOST, RF, XGBOOST, and GBDT methods. The classified outcomes are also labeled as 1 and 0 based on whether the anticipated crime materialized. The findings revealed that the XGBOOST model yielded the highest accuracy, with a prediction rate exceeding 0.7 across all six categories. Subsequently, this predicted data was used as input features for regression analysis employing CNN and LSTM techniques. The results indicate that CNN outperformed the other methods in forecasting the number of crimes across the six categories, achieving MAE values around or below 0.3. This paper presents a two-step machine learning methodology that can effectively forecast future crimes in Chicago, thereby enhancing public safety and enabling the government to uphold law and order.