🚦 Road Accident Risk Prediction: Physics-Informed AI

📋 Project Overview

This project aims to predict the risk of traffic accidents (0-1 probability) based on telematic data. Unlike traditional models that only use raw data, this solution incorporates Physics-Informed Feature Engineering to capture real-world driving dynamics (e.g., centrifugal force) and utilizes a Stacking Ensemble architecture for maximum robustness.

---

🧠 The Strategy: "Physics Meets Data Science"

1. Feature Engineering (Domain Knowledge)

We didn't just feed the model raw numbers. We created synthetic features based on road safety physics:

• Centrifugal Force Proxy: Calculated as Speed Limit × Curvature. High speed on a sharp curve exponentially increases risk.
• Visibility Hazard: Interaction between Lighting (Night/Dim) and Speed.
• Accident Density: Historical accident frequency relative to the road's speed limit.

2. Model Architecture (Stacking Ensemble)

We used a 2-Level Stacking approach to minimize variance and bias:

• Level 0 (Base Learners):
  • 🚀 LightGBM: Optimized for speed and gradient boosting efficiency.
  • 🛡️ XGBoost: Provides stability and regularization.
  • 🐈 CatBoost: Specifically tuned for handling categorical features (Road Type, Weather).
• Level 1 (Meta-Learner):
  • 🧠 Ridge Regression: Learns the optimal weights from the base learners' predictions to generate the final output.

---

🛠️ Tech Stack

• Data Processing: Pandas, NumPy
• Machine Learning: Scikit-Learn (StackingRegressor, RidgeCV)
• Boosting Libraries: XGBoost, LightGBM, CatBoost
• Evaluation Metric: Mean Absolute Error (MAE)

---

📊 Key Findings

• Speed is not the only enemy: Speed alone has a moderate correlation, but Speed combined with Curvature is the highest predictor of accident risk.
• The "Dark" Factor: Driving at night significantly amplifies the risk of other factors (e.g., wet roads), captured successfully by the model's interaction terms.
• Stacking Superiority: The Stacking model outperformed individual models (XGB/LGBM) by effectively blending their strengths.

---

🚀 How to Run

1. Clone the repository:

   git clone [https://github.com/YourUsername/AI-Road-Safety-Risk-Prediction.git](https://github.com/YourUsername/AI-Road-Safety-Risk-Prediction.git)

2. Install dependencies

    pip install pandas xgboost lightgbm catboost scikit-learn

3. Run the notebook

   Road_Accident_Prediction.ipynb

4. Dataset The dataset used in this project is too large for GitHub. You can download it directly from the source: (https://www.kaggle.com/competitions/dstc-kaggle-practice-2025)