🚦 Traffic Accident Type Classification Using Neural Networks

📌 Overview

This project aims to predict the type of traffic accident using a supervised Neural Network (NN) model. The dataset contains nearly 60,000 accident records with driver, vehicle, road, environmental, temporal, and location-related features.

The primary goal is to build a generalizable classification model by applying proper preprocessing, feature engineering, and validation techniques while avoiding data leakage.

---

📊 Dataset

• Size: 59,981 rows × 36 columns

• Target Variable: Accident Type

• Feature Types:

  • Categorical (driver behavior, road condition, weather, vehicle type)
  • Numerical (driver age, vehicle service year, number of vehicles)
  • Temporal (hour, day, month)
  • Spatial (location easting & northing)

---

🔍 Exploratory Data Analysis (EDA)

The following analyses were performed:

• Histograms and boxplots for numerical feature distributions
• Count plots for categorical variables
• Target class distribution analysis
• Correlation analysis for numerical features

EDA helped identify skewness, outliers, and class imbalance.

---

🛠 Data Preprocessing & Feature Engineering

✅ Data Leakage Prevention

Outcome-related variables (e.g., number of casualties, injuries, fatalities) were removed to prevent the model from learning information unavailable before an accident occurs.

✅ Feature Transformations

• Temporal Encoding

  • Hour encoded using sine and cosine to preserve cyclic behavior
  • Additional Time_Bucket feature created (Morning Rush, Midday, Evening Rush, Night)

• Spatial Encoding

  • Location Easting and Location Northing discretized into bins

• Vehicle Age Grouping

  • Vehicles categorized as New, Mid, or Old

• Date Features

  • Date and Year removed to avoid redundancy and leakage

✅ Encoding

• Target Variable: Label encoded
• Categorical Features: Encoded appropriately for NN input
• Numerical Features: Scaled for stable NN training

---

🧠 Model Architecture

• Feedforward Neural Network
• Fully connected dense layers
• ReLU activation
• Softmax output layer for multi-class classification

Loss & Optimization

• Loss Function: Sparse Categorical Crossentropy
• Optimizer: Adam
• Regularization: L2, Dropout, and Early Stopping
• Validation: Used to monitor overfitting and tune hyperparameters

---

📈 Model Evaluation

• Training vs validation loss and accuracy monitoring
• Early stopping to prevent overfitting
• Performance evaluated on unseen data

---

🧪 Key Concepts Applied

• Overfitting prevention
• Cyclic feature encoding
• Label encoding vs one-hot encoding
• Validation-based model tuning
• Data leakage handling

---

🧾 Technologies Used

• Python
• Pandas, NumPy
• Matplotlib, Seaborn
• Scikit-learn
• TensorFlow / Keras
• Jupyter Notebook

---

📌 Future Improvements

• Handle class imbalance using class weights or focal loss
• Try alternative models (XGBoost, Random Forest)
• Add SHAP or feature importance analysis
• Deploy model using a simple web interface

---

👤 Author

Yared Mengiste Software Engineering Student Neural Network & Data Analysis Project