Driving Behaviour Classification Model 🚗

A supervised machine learning project that classifies driver behavior — Safe, Distracted, or Aggressive — using telemetry-style driving data such as speed, braking, steering, and reaction time.

📋 Overview

This Driving Behaviour Machine Learning Based Model trains and compares multiple classification models on a driving telemetry dataset to predict a driver's behavior category from sensor-like inputs. The project covers the full ML workflow: exploratory data analysis, preprocessing, baseline model comparison, cross-validation, hyperparameter tuning, and feature importance analysis.

📂 Dataset

File: Driver_Behavior_Realistic.csv Shape: 29,002 rows × 11 columns

Feature	Description
speed_kmph	Vehicle speed
accel_x, accel_y	Acceleration components
brake_pressure	Braking force applied
steering_angle	Steering wheel angle
throttle	Throttle input
lane_deviation	Deviation from lane center
phone_usage	Phone usage indicator
headway_distance	Distance to vehicle ahead
reaction_time	Driver reaction time
behavior_label	Target — Safe, Distracted, or Aggressive

Class distribution (fairly balanced):

• Distracted: 10,021
• Safe: 10,003
• Aggressive: 8,978

🛠️ Tech Stack

Category	Tools/Tech
Language	Python (Google Colab, T4 GPU runtime)
Data handling	pandas, NumPy
Visualization	matplotlib, seaborn
ML framework	scikit-learn
Models tested	Decision Tree, Random Forest, Gradient Boosting, SVM, KNN, Logistic Regression, MLP (Neural Network)

🧠 Methodology

1. EDA — inspected class balance, feature distributions, and boxplots comparing key features (speed_kmph, brake_pressure, steering_angle, reaction_time) across behavior classes.
2. Preprocessing — label-encoded the target (Aggressive→0, Distracted→1, Safe→2) and standardized features with StandardScaler inside an sklearn Pipeline.
3. Baseline modeling — trained 5 baseline classifiers (Decision Tree, Random Forest, Gradient Boosting, SVM, KNN) and compared Accuracy, Precision, Recall, and F1-Score, with confusion matrices for each.
4. Cross-validation — used Stratified 5-Fold CV (scoring='f1_weighted') for a more robust estimate of each model's performance.
5. Hyperparameter tuning — tuned the top model(s) via GridSearchCV / RandomizedSearchCV (5-fold CV, optimizing weighted F1).
6. Model interpretation — extracted feature importances from Random Forest and Gradient Boosting, and inspected Logistic Regression coefficients per class to understand which features drive each behavior label.
7. Learning rate analysis — tested Gradient Boosting across multiple learning rates (0.5 → 0.01) to study the stability/generalization tradeoff.
8. Final comparison — benchmarked baseline vs. tuned models side-by-side.

🏆 Results

Model	Accuracy	Precision	Recall	F1-Score
KNN (Baseline)	0.9509	0.9509	0.9509	0.9509
Gradient Boosting (Tuned)	0.9509	0.9509	0.9509	0.9509
KNN (Tuned)	0.9507	0.9507	0.9507	0.9507
Neural Network (MLP)	0.9493	0.9493	0.9493	0.9493

Best model: KNN (Baseline) — ~95.1% accuracy/F1, essentially tied with tuned Gradient Boosting, showing the default KNN already performs near-optimally on this dataset without additional tuning overhead.

🚀 Getting Started

Prerequisites

• Python 3.8+
• Jupyter Notebook or Google Colab

Installation

git clone https://github.com/Apoo3va/driver-behaviour-ai.git
cd driver-behaviour-ai
pip install pandas numpy matplotlib seaborn scikit-learn

Running the Notebook

1. Open Driver_Behavior_Fixed.ipynb in Jupyter or Google Colab
2. Place Driver_Behavior_Realistic.csv in the same directory (or upload to /content/ in Colab)
3. Run all cells top to bottom

📁 Project Structure

driving-behaviour-classification/
├── Driver_Behavior_Fixed.ipynb     # Full EDA + model training + comparison notebook
├── Driver_Behavior_Realistic.csv    # Telemetry dataset
└── README.md

🤝 Contributing

Contributions are welcome! Feel free to open an issue or submit a pull request.

👤 Author

Apoorva Yadav

📄 License

This project is licensed under the MIT License — feel free to use, modify, and distribute with attribution.

📬 Contact

For questions or feedback, please open an issue on this repository.