Multi-class classification predicting stellar object type (Galaxy, Star, QSO) from astronomical measurements.
- Competition: Kaggle Playground Series Season 6 Episode 6
- Public Score: 0.96517 (Balanced Accuracy)
- Leaderboard Rank: 1203 / 2,666 teams (Top 45%)
- Submissions: 1
- Feature engineering: color indices (u-g, g-r, r-i, i-z, u-z) from photometric filters
- Model: LightGBM with class_weight='balanced'
- Metric: Balanced Accuracy (handles class imbalance across 3 classes)
- 577,347 training rows, 247,435 test rows
- 3 classes: GALAXY (65.4%), QSO (20.3%), STAR (14.3%)
- Key feature: redshift — almost perfectly separates QSOs from stars and galaxies
| Feature | Description |
|---|---|
| redshift | Most powerful feature — separates QSO from GALAXY and STAR |
| u_g, g_r, r_i, i_z, u_z | Color indices — brightness differences between wavelength filters |
| spectral_type | Star spectrum classification |
| galaxy_population | Galaxy population type |
stellarclass_model/
├── config/
├── processing/
├── trained_models/
├── pipeline.py
train_pipeline.py
predict.py
research.ipynb