Skip to main content

Gradient Boosting (GB)

A scikit-learn Gradient Boosting classifier trained on KOI features to predict exoplanet candidacy with calibrated probability outputs and extensive hyperparameter tuning.

Dataset & target​

  • Data: KOI Playground train/test CSVs with astronomical measurements
  • Target classes: candidate vs non-candidate
  • Features include: koi_period, koi_prad, koi_teq, koi_insol, koi_dor, ... (full list in feature_names.csv)

Pipeline​

  1. Explore and preprocess data (median imputation for numeric)
  2. Encode target variable
  3. Train baseline GB model
  4. GridSearchCV for hyperparameter tuning (primary metric: ROC-AUC)
  5. Final training and evaluation; export artifacts

Gradient Boosting is a reliable “tabular data” model. It learns from many small decision trees, each fixing the mistakes of the previous one.

What it does​

  • Reads KOI table features (period, radius, temperature, etc.)
  • Predicts candidate vs non-candidate
  • Gives a probability for confidence

When to use it​

  • You have a clean table (CSV) with the KOI features
  • You prefer classic ML with quick training and simple deployment

Data needed​

  • Training and test CSVs (KOI features)
  • Missing numeric values filled by median (simple and robust)

How it works (simple)​

  1. Start with a simple tree
  2. Add small trees that focus on remaining errors
  3. Combine all trees to make a strong predictor

Tuning (common knobs)​

  • n_estimators (number of trees), learning_rate, max_depth, subsample
  • Tune with cross-validation, using ROC-AUC or Accuracy

What you’ll see (example visuals)​

Training data

Model visualization

Test set visualizations

Outputs (files)​

  • Model: exoplanet_gradient_boosting_model.joblib
  • Label encoder: label_encoder.joblib
  • Features list: feature_names.csv
  • Predictions and metrics CSVs for reporting

Tips​

  • Match feature order to feature_names.csv
  • Use the same preprocessing at training and prediction time