🚆 Transport Mode Choice Prediction Using Classification Models

📚 Course Information

• Course: AH2179
• Module: 3
• Project: Classification Models for Travel Mode Choice Prediction

🎯 Project Overview

This project implements and compares multiple classification models to predict transportation mode choices (air, bus, rail, car) using various machine learning techniques, with a focus on handling class imbalance.

📊 Data Distribution

• Rail: 41.8% (1288 samples)
• Car: 31.8% (978 samples)
• Air: 22.9% (705 samples)
• Bus: 3.5% (109 samples)

🏗️ Model Architecture

Data Preprocessing

• Features: All columns except 'choice'
• Scaling: StandardScaler implementation
• Encoding: One-hot encoding
• Data Split: 75% training, 25% testing
• Cross-Validation: Stratified K-fold

🤖 Models Implemented

1. Logistic Regression (Best Performer)
2. XGBoost Classifier
3. Random Forests
4. Decision Trees
5. KNN
6. SVM
7. Nearest Centroid

⭐ Best Model Configuration

Logistic Regression parameters:

• C: 0.1
• penalty: 'l2'
• solver: 'newton-cg'

📈 Performance Metrics

Logistic Regression Results

• Accuracy: 57%
• Precision: [0.546, 0.0, 0.471, 0.663]
• Recall: [0.524, 0.0, 0.533, 0.669]

💡 Key Findings & Detailed Reflections

🔍 Model Performance Analysis

1. Class Imbalance Impact

   • Models showed bias towards majority classes (rail, car)
   • KNN and Nearest Centroid uniquely detected 'bus' class despite low overall accuracy
   • Random Forests and Decision Trees showed high sensitivity to imbalanced data

2. Model-Specific Observations

   • Logistic Regression:
     • L2 regularization prevented overfitting
     • Newton-CG solver optimized classification accuracy
   • XGBoost: Performed similarly to Logistic Regression
   • KNN & Nearest Centroid: Better at minority class detection
   • SVM: Struggled with imbalanced classes

🎓 Technical Lessons Learned

1. Data Preprocessing Insights

   • Critical importance of feature encoding methods
   • Impact of training/testing split ratios
   • Value of standardization in model performance

2. Model Selection Considerations

   • LazyPredict library's utility in initial model screening
   • Importance of class_weight parameter
   • Trade-offs between accuracy and minority class detection

3. Optimization Techniques

   • Grid search effectiveness for parameter tuning
   • Cross-validation strategies for imbalanced data
   • Regularization impact on model robustness

🔮 Future Applications & Transport Problems

📋 Potential Applications

1. Bus Delay Classification

   • Handling imbalanced on-time vs delayed data
   • Implementing resampling techniques (SMOTE, oversampling)
   • Ensemble methods for improved performance

2. Traffic Pattern Classification

   • Multi-class prediction for congestion levels
   • Real-time pattern recognition
   • Seasonal variation analysis

3. Infrastructure Usage Prediction

   • Peak vs off-peak utilization
   • Facility type selection
   • Maintenance scheduling

🛠️ Recommended Techniques

1. Handling Imbalanced Data

   • SMOTE (Synthetic Minority Over-sampling Technique)
   • Adaptive synthetic sampling
   • Hybrid sampling approaches

2. Ensemble Methods

   • Voting Classifiers
   • Stacking multiple models
   • Boosting algorithms

3. Advanced Feature Engineering

   • Temporal feature extraction
   • Spatial correlation analysis
   • Domain-specific feature creation

🛠️ Dependencies

• Scikit-learn
• XGBoost
• Pandas
• NumPy
• LazyPredict
• Matplotlib
• Seaborn

---

Note: This project was completed as part of the AH2179 course, Module 3. 🎓