An end-to-end machine learning system for exoplanet detection using NASA astronomical datasets, designed with a strong emphasis on scientific validity, recall-oriented evaluation, and robust generalization.
The project leverages LightGBM along with a custom refit strategy (FitSearchCV) to avoid overfitting and select stable models suitable for real-world scientific discovery.
Astronomical missions generate massive volumes of telescope data, making manual exoplanet detection slow and impractical at scale.
Objective:
Develop an automated and reliable ML pipeline to classify celestial objects into:
- Confirmed Exoplanet
- Exoplanet Candidate
- False Positive
with special focus on minimizing missed exoplanet candidates.
- 🔬 Recall-first evaluation strategy for candidate discovery
- 🧠 LightGBM-based classification for structured astronomical data
- 🛡️ Explicit handling of data leakage and feature integrity
- ⚙️ FitSearchCV: custom hyperparameter refit strategy
- 🌐 Deployed web application for prediction and retraining
- 📊 Careful dataset selection based on label stability
- Clean and well-labeled
- Stable target definitions
- Used for training and testing
- High reported accuracy
- Susceptible to data leakage
- Used only for supplementary validation
- Large and evolving dataset
- Labels under continuous revision
- Not used for supervised training
- Final feature set: 17 features
- Preprocessing steps:
- Removed empty and redundant columns
- Eliminated single-value categorical features
- Excluded target-leaking flags
- False-positive indicators deliberately excluded to preserve model integrity
- Algorithm: LightGBM (Gradient Boosted Decision Trees)
- Why LightGBM?
- Efficient on large tabular datasets
- Handles non-linear feature interactions
- High recall for exoplanet candidates
- Primary Metric: Recall (Candidate class)
Instead of refitting the model with the highest validation score, we optimize for stability using:
min ( |train_score - test_score| + (1 - test_score) ) / 2
- Reduces overfitting
- Avoids underfitting
- Encourages generalizable models
- Better suited for scientific discovery tasks
- 🔍 Real-time single prediction
- 📁 Batch prediction with downloadable results
- 🔄 Online retraining with new datasets
- 🧪 Interactive web interface
- Automated periodic retraining
- Intelligent false-positive flagging
- Dataset validation pipeline
- Continuous monitoring for model drift