Toyproject-ML

1. Predict KBO games using mlp

Dataset:

• Merged 3 seasons (2023–2025) of KBO data (~2,135 games)
• Merged 2 seasons (2024.June–2025) of KBO data (~1000 games)
• csv/: Sliced 2025 data files categorized by batter, pitcher, and scoreboard records from library kbodata

Features (12 inputs):

• Home/Away team’s last 30-game avg runs scored/allowed.
• Home/Away team’s last 30-game batting average.
• Starting pitcher's season avg runs allowed.
• Team's total win % and Home/Away win %.

Model

• Algorithms: Tested SGD, ADAM, and L-BFGS using Scikit-learn and PyTorch.
• Architecture: Settled on a Multilayer Perceptron (MLP) with 3 to 12 hidden nodes and Tanh/ReLU activation.
• Scaling: Applied StandardScaler
• Overfitting Management: Used Dropout and Weight Decay in PyTorch

2. others

Data

• kbo_regular_seasons_result.csv: Historical team performance (1982–2025) including Win/Loss records, BA, ERA, and WPCT from Kaggle

• postseasons_result.csv: Yearly postseason placements and rankings by Team ID.

• data/raw/: Comprehensive historical batting and pitching statistics (1982–2025)

• year_data_2025.json: Detailed game-level data for 2025 (scoreboards, play-by-play events, umpires, audience metrics)

• data/ranked/: Processed datasets containing expanded ranking features.

• Top-level CSVs: Refined datasets prepared for immediate modeling.

Other Notebooks

• ml.ipynb & TabNet_SW.ipynb: Machine learning models for ranking and postseason prediction,TabNet, Time Series
• Rank_process.ipynb: Logic for rank processing and data transformation
• test.ipynb: data extracting from kbodata library