1. Ética y Gobernanza de IA

import pandas as pd

# Verificar balance de clases por grupo protegido
df.groupby(['protected_attribute', 'target']).size()

# Ejemplo: Dataset de crédito
#                    approved
# gender  age_group          
# female  <30           1200
#         50         3400
#         >50           2100
# male    <30           1800
#         50         4200
#         >50           2500

# ⚠️ Warning: Mujeres <30 subrepresentadas

from sklearn.inspection import permutation_importance

# Calcular feature importance por grupo
for group in df['protected_attribute'].unique():
    group_data = df[df['protected_attribute'] == group]
    importance = permutation_importance(model, group_data[features], group_data['target'])
    print(f"Group {group}: {importance.importances_mean}")

# ⚠️ Si features diferentes son importantes para diferentes grupos → posible bias

from imblearn.over_sampling import SMOTE

# Rebalancear dataset por grupo protegido
smote = SMOTE(random_state=42)
X_resampled, y_resampled = smote.fit_resample(X, y)

# Verificar balance
pd.Series(y_resampled).value_counts()

from aif360.algorithms.inprocessing import AdversarialDebiasing

# Entrenar modelo con fairness constraint
debiased_model = AdversarialDebiasing(
    privileged_groups=[{'gender': 1}],  # Hombres
    unprivileged_groups=[{'gender': 0}],  # Mujeres
    scope_name='debiased_classifier',
    debias=True
)
debiased_model.fit(dataset)

P(Y_pred=1 | A=0) = P(Y_pred=1 | A=1)

# Calcular demographic parity
def demographic_parity(y_pred, protected_attr):
    groups = protected_attr.unique()
    rates = {}
    for group in groups:
        mask = protected_attr == group
        rates[group] = y_pred[mask].mean()
    return rates

# Ejemplo: Aprobación de crédito
# Group 0 (mujeres): 0.65
# Group 1 (hombres): 0.70
# Diferencia: 0.05 → ⚠️ Posible bias

P(Y_pred=1 | Y_true=1, A=0) = P(Y_pred=1 | Y_true=1, A=1)  # TPR
P(Y_pred=1 | Y_true=0, A=0) = P(Y_pred=1 | Y_true=0, A=1)  # FPR

from sklearn.metrics import confusion_matrix

def equalized_odds(y_true, y_pred, protected_attr):
    groups = protected_attr.unique()
    metrics = {}
    for group in groups:
        mask = protected_attr == group
        tn, fp, fn, tp = confusion_matrix(y_true[mask], y_pred[mask]).ravel()
        tpr = tp / (tp + fn)
        fpr = fp / (fp + tn)
        metrics[group] = {'TPR': tpr, 'FPR': fpr}
    return metrics

# Ejemplo: Predicción de reincidencia
# Group 0: TPR=0.75, FPR=0.20
# Group 1: TPR=0.80, FPR=0.25
# ⚠️ Diferencias en TPR y FPR → posible bias

P(Y_true=1 | Y_pred=p, A=0) = P(Y_true=1 | Y_pred=p, A=1)

import shap

# Entrenar modelo
model = XGBClassifier()
model.fit(X_train, y_train)

# Calcular SHAP values
explainer = shap.TreeExplainer(model)
shap_values = explainer.shap_values(X_test)

# Visualizar
shap.summary_plot(shap_values, X_test)

from lime.lime_tabular import LimeTabularExplainer

# Crear explainer
explainer = LimeTabularExplainer(
    X_train.values,
    feature_names=X_train.columns,
    class_names=['rejected', 'approved'],
    mode='classification'
)

# Explicar predicción individual
exp = explainer.explain_instance(
    X_test.iloc[0].values,
    model.predict_proba
)
exp.show_in_notebook()

from sklearn.inspection import permutation_importance

# Calcular permutation importance
result = permutation_importance(model, X_test, y_test, n_repeats=10)

# Visualizar
import matplotlib.pyplot as plt
sorted_idx = result.importances_mean.argsort()
plt.barh(X_test.columns[sorted_idx], result.importances_mean[sorted_idx])
plt.xlabel("Permutation Importance")

from diffprivlib.models import LogisticRegression

# Entrenar modelo con differential privacy
model = LogisticRegression(epsilon=1.0)  # Privacy budget
model.fit(X_train, y_train)

# Modelo entrenado sin ver datos individuales directamente

# Pseudocódigo
# Servidor central
global_model = initialize_model()

for round in range(num_rounds):
    # Enviar modelo a clientes
    for client in clients:
        local_model = client.train(global_model, local_data)
        send_to_server(local_model.weights)

    # Agregar pesos locales
    global_model.weights = average(all_local_weights)

# Model Card: Credit Approval Model

## Model Details
- Type: XGBoost Classifier
- Version: 2.1.0
- Trained: 2015

## Intended Use
- Primary use: Approve/reject credit applications
- Out-of-scope: Not for use in countries with different regulations

## Factors
- Evaluated across gender, age, income brackets
- Performance varies by age group (lower accuracy for <25)

## Metrics
- Overall Accuracy: 0.87
- Demographic Parity Difference: 0.03
- Equalized Odds Difference: 0.05

## Training Data
- Source: Internal credit applications 202023
- Size: 500K applications
- Known bias: Underrepresentation of applicants <25

## Ethical Considerations
- Risk: May perpetuate historical bias in credit approval
- Mitigation: Applied SMOTE rebalancing, fairness constraints

# Model Card: [Model Name]

## Model Details
- **Type**: [Classification/Regression/etc]
- **Architecture**: [XGBoost/Neural Network/etc]
- **Version**: [1.0.0]
- **Trained**: [YYYY-MM-DD]
- **Owner**: [Team/Person]

## Intended Use
- **Primary use**: [Description]
- **Out-of-scope uses**: [What NOT to use for]
- **Users**: [Who should use this model]

## Factors
- **Groups**: [Demographic groups evaluated]
- **Instrumentation**: [How data was collected]
- **Environment**: [Deployment environment]

## Metrics
- **Overall Accuracy**: [0.XX]
- **Precision/Recall**: [0.XX / 0.XX]
- **Fairness Metrics**:
  - Demographic Parity Difference: [0.XX]
  - Equalized Odds Difference: [0.XX]

## Training Data
- **Source**: [Where data came from]
- **Size**: [N examples]
- **Timeframe**: [YYYY-MM to YYYY-MM]
- **Known Bias**: [Describe bias]

## Ethical Considerations
- **Risks**: [Potential harms]
- **Mitigations**: [What was done to reduce risks]
- **Limitations**: [Known limitations]

## Caveats and Recommendations
- [Caveat 1]
- [Caveat 2]

# Ethical Review: [Project Name]

## Project Overview
- **Description**: [What the AI system does]
- **Stakeholders**: [Who is affected]
- **Impact**: [Expected impact]

## Risk Assessment

### Bias Risk
- **Risk Level**: [Low/Medium/High]
- **Description**: [Potential bias]
- **Mitigation**: [How to address]

### Privacy Risk
- **Risk Level**: [Low/Medium/High]
- **Description**: [Privacy concerns]
- **Mitigation**: [How to protect privacy]

### Safety Risk
- **Risk Level**: [Low/Medium/High]
- **Description**: [Potential harm]
- **Mitigation**: [Safety measures]

### Fairness Risk
- **Risk Level**: [Low/Medium/High]
- **Description**: [Fairness concerns]
- **Mitigation**: [Fairness interventions]

## Decision
- [ ] Approved
- [ ] Approved with conditions: [List conditions]
- [ ] Rejected: [Reason]

## Monitoring Plan
- **Metrics to track**: [Fairness metrics, performance metrics]
- **Frequency**: [Weekly/Monthly]
- **Responsible**: [Team/Person]