NFL Draft Prediction — High-Accuracy GPU Ensemble

This repository contains the advanced machine learning pipeline developed for the GCI World Cup 2026 In-Class Competition hosted by the University of Tokyo.

🚀 Architecture Overview

The solution implements a highly robust, stabilized ensemble designed to handle tabular noise and prevent leaderboard overfitting:

1. Advanced Feature Engineering: Custom mass-to-motion formulas (SpeedScore, ExplosionScore, StrengthScore) along with positional normalization via Z-Score and historical cross-interactions ({col}_x_Year).
2. Robust Categorical Encoding: 5-fold internal Target Encoding paired with native frequency mapping.
3. Validation Strategy: Strict 10-Fold Stratified Cross-Validation to ensure local stability.
4. Base Models: Optimized instances of LightGBM, CatBoost (GPU), and XGBoost (GPU/CUDA).
5. Within-Fold Rank Scaling: Converts raw probabilities to normalized ranks partition by partition to neutralize calibration scale mismatches.
6. Hyperplane Optimization: A SciPy minimize block using Powell's method to dynamically search for the optimal soft-maxed combination weights based on global Out-of-Fold (OOF) ROC AUC.

📁 Repository Structure

├── data_structure.py         # Initial data structure and baseline experimentation
├── model.py                  # Main architecture: High-Accuracy GPU Ensemble (v5.6 Pro)
├── .gitignore                # Cleans repository from local logs, IDE configurations, and datasets
└── README.md                 # Project documentation

🛠️ Requirements & Environment

The code is optimized to run on a cloud environment with a CUDA-capable GPU (such as Google Colab).

The main dependencies are:

• pandas

• numpy

• scikit-learn

• scipy

• lightgbm

• catboost

• xgboost

🏃 How to Run

Ensure the dataset (train.csv and test.csv) is available in the relative path data/

Run the main model script:

Bash
python model.py
The script will automatically execute the 10-fold cross-validation loop, search for the best rank-blending weights using Powell's optimization, and export the optimized predictions to submission_advanced_ensemble.csv.