Spaces:
Sleeping
Sleeping
| import pandas as pd | |
| import numpy as np | |
| from sklearn.metrics import confusion_matrix | |
| class BiasDetector: | |
| def __init__(self, model, sensitive_features=None): | |
| self.model = model | |
| self.sensitive_features = sensitive_features or [] | |
| self.metrics = {} | |
| def analyze_bias(self, X, y, sensitive_feature): | |
| """Analyze bias for a sensitive feature""" | |
| if sensitive_feature not in X.columns: | |
| raise ValueError(f"Sensitive feature {sensitive_feature} not found in data") | |
| # Get unique values for the sensitive feature | |
| unique_values = X[sensitive_feature].unique() | |
| results = {} | |
| for value in unique_values: | |
| # Filter data for this value | |
| mask = X[sensitive_feature] == value | |
| X_group = X[mask] | |
| y_group = y[mask] | |
| # Skip if no data | |
| if len(X_group) == 0: | |
| continue | |
| # Make predictions | |
| if hasattr(self.model, 'predict'): | |
| y_pred = self.model.predict(X_group) | |
| else: | |
| y_pred = self.model.model.predict(X_group) | |
| # Calculate metrics | |
| tn, fp, fn, tp = confusion_matrix(y_group, y_pred).ravel() | |
| # Calculate rates | |
| acceptance_rate = (tp + fp) / len(y_group) | |
| true_positive_rate = tp / (tp + fn) if (tp + fn) > 0 else 0 | |
| false_positive_rate = fp / (fp + tn) if (fp + tn) > 0 else 0 | |
| results[value] = { | |
| "count": len(X_group), | |
| "acceptance_rate": acceptance_rate, | |
| "true_positive_rate": true_positive_rate, | |
| "false_positive_rate": false_positive_rate | |
| } | |
| self.metrics[sensitive_feature] = results | |
| return results | |
| def calculate_disparate_impact(self, sensitive_feature, privileged_value): | |
| """Calculate disparate impact""" | |
| if sensitive_feature not in self.metrics: | |
| raise ValueError(f"No metrics available for {sensitive_feature}. Run analyze_bias first.") | |
| metrics = self.metrics[sensitive_feature] | |
| if privileged_value not in metrics: | |
| raise ValueError(f"Privileged value {privileged_value} not found in metrics") | |
| privileged_rate = metrics[privileged_value]["acceptance_rate"] | |
| disparate_impact = {} | |
| for value, data in metrics.items(): | |
| if value != privileged_value: | |
| unprivileged_rate = data["acceptance_rate"] | |
| # Disparate impact is the ratio of unprivileged to privileged acceptance rates | |
| # A value < 0.8 is often considered problematic | |
| di = unprivileged_rate / privileged_rate if privileged_rate > 0 else 0 | |
| disparate_impact[value] = di | |
| return disparate_impact | |
| def generate_bias_report(self): | |
| """Generate a comprehensive bias report""" | |
| report = {} | |
| for feature in self.metrics: | |
| feature_report = { | |
| "metrics": self.metrics[feature], | |
| "disparate_impact": {} | |
| } | |
| # Find privileged group (highest acceptance rate) | |
| acceptance_rates = {value: data["acceptance_rate"] | |
| for value, data in self.metrics[feature].items()} | |
| privileged_value = max(acceptance_rates, key=acceptance_rates.get) | |
| # Calculate disparate impact | |
| feature_report["disparate_impact"] = self.calculate_disparate_impact( | |
| feature, privileged_value | |
| ) | |
| # Add to report | |
| report[feature] = feature_report | |
| return report |