A machine learning pipeline that classifies Kepler Objects of Interest as confirmed exoplanets or false positives using transit signal and stellar properties.
NASA's Kepler Space Telescope observed ~200,000 stars for four years, detecting periodic brightness dips that could indicate orbiting planets. Not every signal is a real planet — eclipsing binary stars, instrumental noise, and stellar variability produce similar signatures. This project builds a classifier that distinguishes real planets from false positives using 9 physical features extracted from Kepler's observations.
- Accuracy: 91.3%
- ROC AUC: 0.968
- Planet Precision: 86%, when the model says "planet," it's right 86% of the time
- Planet Recall: 86%, the model finds 86% of real planets
- Transit depth (
koi_depth) and orbital period (koi_period) are the strongest predictors, consistent with astrophysical expectations - No single feature cleanly separates the classes — the model's power comes from combining all 9 features simultaneously
- The model's errors concentrate in the moderate transit depth range where planet and eclipsing binary signatures overlap
- SHAP analysis confirms the model learned physically meaningful patterns rather than arbitrary correlations
Kepler Exoplanet Search Results from Kaggle. ~10,000 Kepler Objects of Interest with pre-extracted transit and stellar features.
Download cumulative.csv and place it in the data/ directory.
Python, pandas, scikit-learn, matplotlib, seaborn, SHAP
exoplanet-classifier/
├── data/ # cumulative.csv goes here
├── notebooks/
│ ├── 01_eda.ipynb # Exploratory data analysis
│ ├── 02_preprocessing.ipynb # Feature engineering and pipeline
│ ├── 03_modeling.ipynb # Model training and comparison
│ ├── 04_evaluation.ipynb # Tuning and evaluation metrics
│ └── 05_visualization.ipynb # Hero plot, SHAP, error analysis
├── .gitignore
├── requirements.txt
└── README.md
git clone https://github.com/GitWorkingTime/exoplanet-classifier.git
cd exoplanet-classifier
python -m venv venv
source venv/bin/activate # or venv\Scripts\activate on Windows
pip install -r requirements.txtDownload the dataset from Kaggle and place cumulative.csv in the
data/ folder. Then open the notebooks in order.
- Feature engineering matters more than model selection, log transformations on skewed features improved model performance significantly
- Evaluation metrics beyond accuracy are essential — with imbalanced classes, F1 score and precision-recall curves give a more honest picture than accuracy alone
- SHAP explanations connect machine learning to domain knowledge, verifying that the model uses astrophysically meaningful patterns builds trust in its predictions