🚧 Highway to... Predicting Fatal Traffic Accidents in Brazil

This repository contains the code and instructions for reproducing the results presented in the paper:

Highway to... Determining Fatal Outcomes in Traffic Accidents Based on Police Reports
Submitted to BRACIS 2025
(Authors omitted for review)

Overview

This study investigates the use of machine learning to predict the severity of traffic accidents in southern Brazil using open data from the Brazilian Federal Highway Police (PRF). We compare the performance of Random Forest (RF), k-Nearest Neighbors (kNN), and Multilayer Perceptron (MLP) classifiers, applying preprocessing techniques such as categorical encoding, feature selection, and SMOTE for class balancing.

The best performing models are further analyzed using SHAP (SHapley Additive exPlanations) to provide insight into the most influential features contributing to accident fatality predictions.

📁 Repository Structure

.
├── data/
│ └── README.md # Instructions to download official PRF dataset
├── models/ # Directory for saving or loading pre-trained models
├── results/ # Directory for saving performance metrics and plots
├── train_models.R # Script to preprocess data and train all models
├── explain.R # Script to compute and visualize SHAP values
└── README.md # You're here :)

Data Acquisition

Due to size and licensing constraints, we do not host the raw datasets here. To use the training script, please follow the instructions in data/README.md to download the required data directly from official PRF sources.

👉 Pre-processed datasets are avaialable as well:

Trainning set without SMOTE: https://drive.google.com/file/d/1kKGdiQz6M_kpNIpU4nlsRlFh-vjNolYh/view?usp=share_link

Trainning set with SMOTE: https://drive.google.com/file/d/16eS_T6T_fZ7H5n88wbayhhKMtXVS4_NS/view?usp=share_link

Validation set: https://drive.google.com/file/d/1AWJ7l5aIqSyBo-0EAom9f2xBJO506Dix/view?usp=share_link

Training the Models

To train all models from scratch:

source("train_models.R")

This script performs:

• Data loading and preprocessing
• Class balancing using SMOTE
• Training of RF, kNN, and MLP models
• Evaluation through metrics: F1-score, specificity, and AUC-ROC
• Model persistence in models/ directory

👉 Or download pre-trained models:

Models without SMOTE: https://drive.google.com/file/d/1L_QRj-K4PtuZvwV0vh49DwKcyd3SQBOB/view?usp=share_link

Models with SMOTE: https://drive.google.com/file/d/19QhEcVT6uvseQ9RcftyI0aTzELu5wLOR/view?usp=share_link

Model Explainability

To generate SHAP-based explanations for the RF model:

source("explain.R")

The explainability script includes:

• SHAP value computation
• Global feature importance visualization
• Local explanation trough beeswarm plot
• Categorical features individual factors explanations

👉 Pre-computed SHAP values are available here:

Download SHAP Values: https://drive.google.com/file/d/1WB6KWUGQDfdrP1vIX87BAwpOZji3fnz3/view?usp=share_link

Encoding maps: https://drive.google.com/file/d/1Y_vXTFcKNzhQJzk6gHo6-LvcdUEKd9uM/view?usp=share_link