kevinkyi/Homework2_Classical_ML

Football Elite Classifier — AutoML (AutoGluon Tabular)

Purpose

This model was developed as part of a class assignment on designing and deploying AI/ML systems.
It demonstrates the use of AutoML (AutoGluon Tabular) to build a binary classifier on football receiver stats.

Dataset

• Source: https://huggingface.co/datasets/james-kramer/receiverstats
• Split: Stratified Train/Test = 80/20 on the original split.
• Features: ['Tgt', 'Rec', 'Yds', 'YBC_per_R', 'YAC_per_R', 'ADOT', 'Drop_pct', 'Rat']
• Target: Elite (0/1)
• Preprocessing: Identifier columns dropped (e.g., Player). Numeric coercion applied; rows with NA removed.

Training Setup

• Framework: AutoGluon Tabular
• Preset: best_quality
• Time budget: 300 seconds
• Seed: 42
• Eval metric: F1 (binary)
• Hardware/Compute: Colab CPU runtime (2 vCPUs, ~12 GB RAM)
• AI Usage Disclosure: Generative AI tools were used to help structure code and documentation; model training and results are real.

Hyperparameters / Search Space

• AutoGluon explored LightGBM, XGBoost, and ensembling variants.
• Random state set for reproducibility.
• Auto-stacking and bagging enabled under best_quality.
• Internal hyperparameter tuning handled automatically by AutoGluon.

Results (Held-out Test)

{
  "accuracy": 0.8333333333333334,
  "f1": 0.8
}

Limitations & Ethics

• Correlations do not imply causation; labels may reflect selection bias.
• Out-of-distribution players/contexts may reduce performance.
• Intended for coursework, not for real personnel decisions.

License

• Code & weights: <MIT/Apache-2.0 or course-required license>

Acknowledgments

AutoML with [AutoGluon Tabular]. Trained in Google Colab. GenAI tools assisted with boilerplate and doc structure. James Kramers hugging face dataset