Model Artifacts

1. Random Forest Classifier

• File: random_forest_classifier.pkl
• Type: Scikit-learn RandomForestClassifier
• Size: ~5-10 MB
• Description: Ensemble classifier combining multiple decision trees for robust predictions
• Performance: ~90% accuracy, AUC-ROC 0.94-0.96

2. Decision Tree Classifier

• File: decision_tree_classifier.pkl
• Type: Scikit-learn DecisionTreeClassifier
• Size: ~1-3 MB
• Description: Single decision tree for interpretable rule-based classification
• Performance: ~85% accuracy, AUC-ROC 0.88-0.91

3. CNN Model

• File: cnn_model.h5
• Type: TensorFlow/Keras Sequential neural network
• Size: ~500 KB - 2 MB
• Description: Deep learning model with dense layers and dropout regularization
• Architecture: Input(9) → Dense(64) → Dropout(0.3) → Dense(32) → Dropout(0.3) → Dense(16) → Dropout(0.3) → Output(1)
• Performance: ~89% accuracy, AUC-ROC 0.92-0.95

4. k-Nearest Neighbors

• File: knn_model.pkl
• Type: Scikit-learn KNeighborsClassifier
• Size: ~10-20 MB (stores training data)
• Description: Instance-based classifier using k=5 nearest neighbors
• Performance: ~87% accuracy, AUC-ROC 0.90-0.93

5. Standard Scaler

• File: scaler.pkl
• Type: Scikit-learn StandardScaler
• Size: ~1-5 KB
• Description: Feature normalization to zero mean and unit variance
• Critical: Must use the same scaler for training and inference

6. Simple Imputer

• File: imputer.pkl
• Type: Scikit-learn SimpleImputer
• Size: ~1-5 KB
• Description: Fill missing values using mean strategy
• Critical: Must use the same imputer for training and inference

Model Architectures & Hyperparameters

Random Forest

• n_estimators: 100
• max_depth: None
• threshold: 0.5

CNN

• hidden_layers: [64,32,16]
• dropout: 0.3
• epochs: 50
• threshold: 0.6

kNN

• n_neighbors: 5
• metric: euclidean
• threshold: 0.6

Decision Tree

• max_depth: 10
• criterion: gini
• threshold: 0.6

Performance Comparison

Prediction (auto-load)

Models are loaded automatically on first prediction. The detector checks if the model is loaded and calls the corresponding load method if needed.

from exobengal import DetectExoplanet

detector = DetectExoplanet()
sample = [365.0, 1.0, 288.0, 1.0, 4.44, 5778, 0.1, 5.0, 100.0]
print(detector.random_forest(sample))

Explicit Pre-loading

from exobengal import DetectExoplanet

detector = DetectExoplanet()
detector.load_rf_model()  # Pre-load for faster predictions

# Now predictions are faster
result = detector.random_forest([365.0, 1.0, 288.0, 1.0, 4.44, 5778, 0.1, 5.0, 100.0])
print(result)

Retraining

Each train_* method overwrites its model file and updates scaler.pkl and imputer.pkl. Always backup existing models before retraining.

Important Notes

• Each train_* overwrites its model and writes scaler.pkl and imputer.pkl
• Always backup existing models before retraining: cp -r models/ models_backup/
• Training automatically handles data loading, preprocessing, splitting (80/20), and evaluation
• Outputs include classification report, confusion matrix, and AUC-ROC score

Retraining Example

from exobengal import DetectExoplanet

detector = DetectExoplanet()

# Retrain Random Forest with custom hyperparameters
detector.train_random_forest(
    data_path="data/cumulative.csv",
    n_estimators=200,
    max_depth=20
)

Loading Custom Models

You can specify custom model paths when initializing the detector:

from exobengal import DetectExoplanet

detector = DetectExoplanet(
    rf_model_path="/path/to/my_rf_model.pkl",
    scaler_path="/path/to/my_scaler.pkl",
    imputer_path="/path/to/my_imputer.pkl"
)

Troubleshooting

Common Issues

• Model file not found: Train models first or download from repository
• TensorFlow warnings: Safe to ignore or install tensorflow-cpu
• Inconsistent predictions: Ensure same scaler/imputer used for training and inference