Prediction of Traffic Accident Severity in Chicago: A Comparative Study of Random Forest and Neural Network Models

This project builds a high-severity accident prediction system based on traffic accident data from the Chicago metropolitan area. Its core goal is to provide a fair benchmark for model selection by rigorously comparing the performance of random forest classifiers and feedforward neural networks. The study has important reference value for data science practice and traffic accident management.

Traffic accidents are one of the main causes of casualties and property losses worldwide, with tens of thousands of accidents occurring in Chicago each year. Accurately predicting accident severity is of great significance for optimizing rescue resources, assessing claim risks, and formulating preventive measures. Traditional post-hoc statistics are difficult to evaluate in real time, and machine learning brings new possibilities for data-driven prediction.

Dataset: Uses Chicago's public traffic accident data, including features such as time, location, weather, and road conditions. Feature Engineering: Handles missing values, category encoding, feature selection, and data imbalance issues. Model Comparison: Random Forest (strong interpretability, anti-overfitting) vs. Feedforward Neural Network (nonlinear modeling, automatic feature learning); uses the same data division, evaluation metrics, and feature set, and ensures fairness through hyperparameter tuning and multiple experiments.

The prediction system can be applied to: 1. Emergency response optimization (real-time resource scheduling); 2. Insurance claim assessment (quick amount estimation); 3. Traffic safety research (identifying key risk factors); 4. Intelligent transportation systems (dynamic early warning).

Limitations: Insufficient data timeliness, geographical limitations, and feature integrity issues. Future Directions: Introduce time-series models to capture dynamic patterns, integrate multi-source data (real-time traffic, weather radar), and develop models with stronger interpretability.

This project demonstrates the performance differences between the two models through rigorous comparative experiments, and the fair comparison methodology provides a reference for machine learning practice. As an open-source project, the code and process are publicly reproducible, making it an excellent case for data science learning, covering the complete process from problem definition to result evaluation.