model-explainer（模型解释器）

Model Explainer

Overview

Makes black-box models interpretable. Explains why models make specific predictions, which features matter most, and how features interact. Critical for trust, debugging, and regulatory compliance.

Why Explainability Matters

• Trust: Stakeholders trust models they understand
• Debugging: Find model weaknesses and biases
• Compliance: GDPR, fair lending laws require explanations
• Improvement: Understand what to improve
• Safety: Detect when model might fail

Explanation Types

1. Global Explanations (Model-Level)

Feature Importance:

from specweave import explain_model

explainer = explain_model(
    model=trained_model,
    X_train=X_train,
    increment="0042"
)

# Global feature importance
importance = explainer.feature_importance()

Output:

Top Features (Global):
1. transaction_amount (importance: 0.35)
2. user_history_days (importance: 0.22)
3. merchant_reputation (importance: 0.18)
4. time_since_last_transaction (importance: 0.15)
5. device_type (importance: 0.10)

Partial Dependence Plots:

# How does feature affect prediction?
explainer.partial_dependence(feature="transaction_amount")

2. Local Explanations (Prediction-Level)

SHAP Values:

# Explain single prediction
explanation = explainer.explain_prediction(X_sample)

Output:

Prediction: FRAUD (probability: 0.92)

Why?
+ transaction_amount=5000 → +0.45 (high amount increases fraud risk)
+ user_history_days=2 → +0.30 (new user increases risk)
+ merchant_reputation=low → +0.25 (suspicious merchant)
- time_since_last_transaction=1hr → -0.08 (recent activity normal)

Base prediction: 0.10
Final prediction: 0.92

LIME Explanations:

# Local interpretable model
lime_exp = explainer.lime_explanation(X_sample)

Usage in SpecWeave

from specweave import ModelExplainer

# Create explainer
explainer = ModelExplainer(
    model=model,
    X_train=X_train,
    feature_names=feature_names,
    increment="0042"
)

# Generate all explanations
explainer.generate_all_reports()

# Creates:
# - feature-importance.png
# - shap-summary.png
# - pdp-plots/
# - local-explanations/
# - explainability-report.md

Real-World Examples

Example 1: Fraud Detection

# Explain why transaction flagged as fraud
transaction = {
    "amount": 5000,
    "user_age_days": 2,
    "merchant": "new_merchant_xyz"
}

explanation = explainer.explain(transaction)
print(explanation.to_text())

Output:

FRAUD ALERT (92% confidence)

Main factors:
1. Large transaction amount ($5000) - Very unusual for new users
2. Account only 2 days old - Fraud pattern
3. Merchant has low reputation score - Red flag

If this is legitimate:
- User should verify identity
- Merchant should be manually reviewed

Example 2: Loan Approval

# Explain loan rejection
applicant = {
    "income": 45000,
    "credit_score": 620,
    "debt_ratio": 0.45
}

explanation = explainer.explain(applicant)
print(explanation.to_text())

Output:

LOAN DENIED

Main reasons:
1. Credit score (620) below threshold (650) - Primary factor
2. High debt-to-income ratio (45%) - Risk indicator
3. Income ($45k) adequate but not strong

To improve approval chances:
- Increase credit score by 30+ points
- Reduce debt-to-income ratio below 40%

Regulatory Compliance

GDPR "Right to Explanation"

# Generate GDPR-compliant explanation
gdpr_explanation = explainer.gdpr_explanation(prediction)

# Includes:
# - Decision rationale
# - Data used
# - How to contest decision
# - Impact of features

Fair Lending Act

# Check for bias in protected attributes
bias_report = explainer.fairness_report(
    sensitive_features=["gender", "race", "age"]
)

# Detects:
# - Disparate impact
# - Feature bias
# - Recommendations for fairness

Visualization Types

1. Feature Importance Bar Chart
2. SHAP Summary Plot (beeswarm)
3. SHAP Waterfall (single prediction)
4. Partial Dependence Plots
5. Individual Conditional Expectation (ICE)
6. Force Plots (interactive)
7. Decision Trees (surrogate models)

Integration with SpecWeave

# Generate all explainability artifacts
/ml:explain-model 0042

# Explain specific prediction
/ml:explain-prediction --increment 0042 --sample sample.json

# Check for bias
/ml:fairness-check 0042

Explainability artifacts automatically included in increment documentation and COMPLETION-SUMMARY.

Best Practices

1. Generate explanations for all production models - No "black boxes" in production
2. Check for bias - Test sensitive attributes
3. Document limitations - What model can't explain
4. Validate explanations - Do they make domain sense?
5. Make explanations accessible - Non-technical stakeholders should understand

Model explainability is non-negotiable for responsible AI deployment.