bluealert-api / ml_model /explainability.py
VandnaShukla-15
Initial BlueAlert backend deployment
63399ee
Raw History Blame Contribute Delete
3.67 kB
"""Model explainability module using feature importance for Blue Alert predictions."""
import numpy as np
import json
class ModelExplainer:
"""Generates feature importance-based explanations for water quality predictions."""
def __init__(self, model, scaler=None):
self.model = model
self.scaler = scaler
self.feature_names = [
'ph', 'Hardness', 'Solids', 'Chloramines', 'Sulfate',
'Conductivity', 'Organic_carbon', 'Trihalomethanes', 'Turbidity',
]
def get_feature_importance(self):
"""Get feature importance from the model."""
if hasattr(self.model, 'feature_importances_'):
importances = self.model.feature_importances_
return dict(zip(self.feature_names, importances.tolist()))
if hasattr(self.model, 'estimator'):
base = self.model.estimator
if hasattr(base, 'feature_importances_'):
importances = base.feature_importances_
return dict(zip(self.feature_names, importances.tolist()))
if hasattr(self.model, 'calibrated_classifiers_'):
for cc in self.model.calibrated_classifiers_:
base = cc.estimator
if hasattr(base, 'feature_importances_'):
importances = base.feature_importances_
return dict(zip(self.feature_names, importances.tolist()))
return None
def _get_tree_based_estimator(self):
"""Navigate through model wrappers to find a tree-based estimator."""
base_model = self.model
if hasattr(self.model, 'calibrated_classifiers_'):
base_model = self.model.calibrated_classifiers_[0].estimator
if hasattr(base_model, 'named_estimators_'):
for est_name, est in base_model.named_estimators_.items():
if hasattr(est, 'feature_importances_'):
return est
if hasattr(base_model, 'feature_importances_'):
return base_model
return None
def explain_prediction(self, features, feature_names=None):
"""Explain a single prediction using feature importance.
SHAP is skipped entirely due to numexpr compatibility issues.
Feature importance provides sufficient model transparency for water quality predictions.
"""
if feature_names is None:
feature_names = self.feature_names
features_array = np.array(features).reshape(1, -1)
try:
# Get feature importance from model
importance = self.get_feature_importance()
if importance:
# Rank features by importance
impact_ranking = sorted(
importance.items(),
key=lambda x: abs(x[1]),
reverse=True,
)
return {
'method': 'FeatureImportance',
'feature_importance': importance,
'features': dict(zip(feature_names, features.tolist() if hasattr(features, 'tolist') else features)),
'impact_ranking': [
{'feature': f, 'importance': round(imp, 4)}
for f, imp in impact_ranking
],
'note': 'Using feature importance for model transparency',
}
else:
return {
'error': 'No feature importance available from model',
}
except Exception as e:
return {
'error': f'Explanation generation failed: {str(e)}',
}