shap
Explain model predictions with SHAP
Also available from: davila7
Machine learning models often act as black boxes. SHAP provides a unified framework to explain any model prediction by computing feature contributions using Shapley values from game theory. Use this skill to visualize feature importance, debug model behavior, and implement explainable AI.
Download the skill ZIP
Upload in Claude
Go to Settings → Capabilities → Skills → Upload skill
Toggle on and start using
Test it
Using "shap". Explain my XGBoost model's predictions for customer churn
Expected outcome:
- Top drivers of churn: Contract Type (0.25), Monthly Charges (0.18), Tenure (0.15)
- Customers without annual contracts are 25% more likely to churn
- Higher monthly charges increase churn probability when tenure is low
- Loyal customers (36+ months) have 60% lower churn risk
Using "shap". Why was this loan application denied
Expected outcome:
- Key factors reducing approval probability: High debt-to-income ratio (-0.22), Short employment history (-0.15), Low credit utilization ratio (-0.12)
- Positive factors: Stable income (+0.08), Good payment history (+0.05)
- Missing factors: No recent credit inquiries helped slightly (+0.03)
Security Audit
SafeDocumentation-only skill with no executable code. All 295 static findings are false positives: markdown code blocks flagged as shell backticks, 'SHAP' acronym flagged as SHA cryptographic algorithm, GitHub URL path component flagged as C2, and standard documentation URLs. This is a pure reference guide for the SHAP machine learning library.
Risk Factors
⚙️ External commands (5)
🌐 Network access (2)
Quality Score
What You Can Build
Debug model behavior
Identify why models make errors by examining SHAP values for misclassified samples and detecting data leakage.
Add explanations to APIs
Integrate SHAP explanations into production prediction services with feature attribution for each prediction.
Validate model fairness
Check for biased predictions across demographic groups by comparing SHAP values for protected attributes.
Try These Prompts
Generate a SHAP beeswarm plot to show feature importance for my XGBoost model trained on my tabular dataset. Show the top 15 most important features with their value distributions.
Explain why my model predicted a specific outcome for a single sample. Generate a waterfall plot showing how each feature contributed to the prediction.
Compare feature importance between my Random Forest and XGBoost models using SHAP. Show which features are ranked differently between the two models.
Analyze my model for potential bias across gender and age groups. Compare SHAP values for the protected attributes and identify if any proxies exist.
Best Practices
- Use TreeExplainer for tree-based models instead of KernelExplainer for faster computation
- Start with global plots (beeswarm, bar) to understand overall patterns before examining individual predictions
- Select 50-1000 representative background samples for accurate SHAP value computation
Avoid
- Using KernelExplainer for tree models when TreeExplainer is available (much slower)
- Interpreting SHAP values as causal relationships without domain validation
- Ignoring model output type (log-odds vs probability) when interpreting SHAP values