A sophisticated machine learning system designed to forecast boxing match outcomes with high precision. The project leverages both deep learning architectures (Multi-Layer Perceptron - MLP) and state-of-the-art gradient boosting techniques (LightGBM) to analyze comprehensive boxer statistics and historical match data. The system employs robust validation methods including k-fold cross-validation to ensure reliable and generalizable predictions.

The system combines neural networks and gradient boosting for optimal performance, implementing 5-fold stratified cross-validation for robust model evaluation. It processes extensive boxer statistics and historical match data through a flexible inference system that supports multiple model types and custom data inputs, complete with detailed performance metrics and training progress visualization.

• Source Codehttps://github.com/axortiz/DL_Final_Project
• Technology StackPyTorch, LightGBM, NumPy, Pandas, Scikit-learn, Matplotlib
• Key Features
  • Dual Model Architecture: Combining neural networks and gradient boosting for enhanced prediction accuracy
  • Advanced Validation: Implementing 5-fold stratified cross-validation for robust model evaluation
  • Data Analysis: Processing comprehensive boxer statistics and historical match data for accurate predictions
  • Flexible Inference: Supporting multiple model types with custom data inputs and configurable parameters
  • Rich Visualization: Detailed performance metrics and training progress visualization tools
• Technical Achievements
  • Data Pipeline: Automated normalization and preprocessing for efficient data handling
  • Training Optimization: Dynamic learning rate adjustment with adaptive scheduling
  • Model Management: Sophisticated checkpoint system for optimal model preservation
  • Performance: GPU-accelerated training with LightGBM for high-speed processing
  • Robustness: Advanced regularization techniques for reliable predictions

Model Architecture

Neural Network Architecture

Training and Validation Accuracy Over Time

Training and Validation Loss Curves

5-Fold Cross-Validation Performance