DarwinDanish/exoplanet-classifier-stacking

Model Card: Exoplanet Candidate Classifier (Stacking Ensemble)

Model Details

Model Description

This is a robust machine learning pipeline designed to classify Kepler Objects of Interest (KOIs). It determines whether a detected signal represents a real exoplanet or a false positive.

The model utilizes a Stacking Ensemble architecture, combining the predictions of three powerful gradient boosting frameworks (LightGBM, XGBoost, and CatBoost) and aggregating them using a final LightGBM meta-learner. It is specifically engineered to handle missing data (NaNs) in scientific datasets through a dual-strategy imputation pipeline.

• Developed by: [Darwin Danish]
• Model Type: Scikit-learn Pipeline (StackingClassifier)
• Input: Tabular data (16 astrophysical features)
• Output: Multi-class classification (CANDIDATE, CONFIRMED, FALSE POSITIVE)

Model Sources

• Repository: https://huggingface.co/DarwinDanish/exoplanet-classifier-stacking
• Dataset Source: NASA Kepler Object of Interest (KOI) Table

---

Uses

Direct Use

This model is intended for astronomers, data scientists, and space enthusiasts who want to analyze Kepler mission data or similar photometric datasets. It predicts the "disposition" of a celestial object based on its physical properties.

Supported Features (Input)

To use this model, your input DataFrame must contain the following columns:

Critical Features:

• koi_period: Orbital period
• koi_depth: Transit depth
• koi_prad: Planetary radius
• koi_sma: Semi-major axis
• koi_teq: Equilibrium temperature
• koi_insol: Insolation flux
• koi_model_snr: Signal-to-Noise Ratio

Auxiliary Features:

• koi_time0bk, koi_duration, koi_incl, koi_srho, koi_srad, koi_smass, koi_steff, koi_slogg, koi_smet

---

How to Get Started with the Model

You can load this model directly from the Hugging Face Hub using joblib and huggingface_hub.

1. Installation

pip install huggingface_hub joblib pandas scikit-learn lightgbm xgboost catboost

2. Python Inference Code

import joblib
import pandas as pd
from huggingface_hub import hf_hub_download

# 1. Download the model and label encoder
repo_id = "DarwinDanish/exoplanet-classifier-stacking"

model_path = hf_hub_download(repo_id=repo_id, filename="exo_stacking_pipeline.pkl")
encoder_path = hf_hub_download(repo_id=repo_id, filename="exo_label_encoder.pkl")

# 2. Load the artifacts
pipeline = joblib.load(model_path)
label_encoder = joblib.load(encoder_path)

# 3. Create sample data (Example: A likely planet candidate)
# Note: The model handles NaNs, so missing values are allowed.
data = {
    'koi_period': [365.25],
    'koi_depth': [1000.5],
    'koi_prad': [1.02],   # Earth radii
    'koi_sma': [1.0],     # AU
    'koi_teq': [255.0],   # Kelvin
    'koi_insol': [1.0],
    'koi_model_snr': [35.5],
    # Aux features (can be mostly defaults or NaNs)
    'koi_time0bk': [135.0],
    'koi_duration': [4.5],
    'koi_incl': [89.9],
    'koi_srho': [1.0],
    'koi_srad': [1.0],
    'koi_smass': [1.0],
    'koi_steff': [5700],
    'koi_slogg': [4.5],
    'koi_smet': [0.0]
}

df_new = pd.DataFrame(data)

# 4. Predict
prediction_index = pipeline.predict(df_new)
prediction_label = label_encoder.inverse_transform(prediction_index)
probabilities = pipeline.predict_proba(df_new)

print(f"Prediction: {prediction_label[0]}")
print(f"Confidence: {max(probabilities[0]):.4f}")

---

Training Details

Training Procedure

The model was trained using a robust preprocessing pipeline followed by a Stacking Classifier.

1. Preprocessing

The pipeline splits features into two groups with different imputation strategies:

• Critical Features: Missing values filled with constant -999. Scaled via StandardScaler.
• Auxiliary Features: Missing values filled with the median. Scaled via StandardScaler.

2. Architecture

• Level 0 (Base Learners):
  • LightGBM: (500 estimators, GPU accelerated)
  • XGBoost: (500 estimators, Histogram tree method, GPU accelerated)
  • CatBoost: (500 estimators, Depth 8, GPU accelerated)
• Level 1 (Meta Learner):
  • LightGBM: (200 estimators) - Aggregates the probabilities from Level 0 to make the final decision.

Feature Importance

Based on the base learners, the most critical features for classification were identified as:

1. koi_model_snr (Signal-to-Noise Ratio)
2. koi_prad (Planetary Radius)
3. koi_depth (Transit Depth)
4. koi_period (Orbital Period)

---

Evaluation Results

The model was evaluated on a held-out test set (20% of data) using stratified splitting.

• Accuracy: ~90%+ (Dependent on specific test split)
• Precision/Recall: High precision in distinguishing False Positives from Candidates.

---

Bias, Risks, and Limitations

• Data Specificity: This model is trained specifically on Kepler mission data. It may not generalize well to data from TESS or JWST without fine-tuning, as the instrumentation and noise profiles differ.
• Class Imbalance: Depending on the dataset version, "False Positives" are often more numerous than "Confirmed" planets, which can bias the model slightly toward false positive predictions in low-SNR ranges.

Environmental Impact

• Compute: Trained on GPU (NVIDIA T4/P100 class) via Kaggle Kernels.
• Training Time: < 5 minutes due to GPU acceleration and efficient gradient boosting implementations.

---