shap（SHAP 模型解释）

SHAP Model Explainability

Explain ML predictions using Shapley values - feature importance and attribution.

When to Use

• Explain why a model made specific predictions
• Calculate feature importance with attribution
• Debug model behavior and validate predictions
• Create interpretability plots (waterfall, beeswarm, bar)
• Analyze model fairness and bias

Quick Start

import shap
import xgboost as xgb

# Train model
model = xgb.XGBClassifier().fit(X_train, y_train)

# Create explainer
explainer = shap.TreeExplainer(model)

# Compute SHAP values
shap_values = explainer(X_test)

# Visualize
shap.plots.beeswarm(shap_values)

Choose Explainer

# Tree-based models (XGBoost, LightGBM, RF) - FAST
explainer = shap.TreeExplainer(model)

# Deep learning (TensorFlow, PyTorch)
explainer = shap.DeepExplainer(model, background_data)

# Linear models
explainer = shap.LinearExplainer(model, X_train)

# Any model (slower but universal)
explainer = shap.KernelExplainer(model.predict, X_train[:100])

# Auto-select best explainer
explainer = shap.Explainer(model)

Compute SHAP Values

# Compute for test set
shap_values = explainer(X_test)

# Access components
shap_values.values      # SHAP values (feature attributions)
shap_values.base_values # Expected model output (baseline)
shap_values.data        # Original feature values

Visualizations

Global Feature Importance

# Beeswarm - shows distribution and importance
shap.plots.beeswarm(shap_values)

# Bar - clean summary
shap.plots.bar(shap_values)

Individual Predictions

# Waterfall - breakdown of single prediction
shap.plots.waterfall(shap_values[0])

# Force - additive visualization
shap.plots.force(shap_values[0])

Feature Relationships

# Scatter - feature vs SHAP value
shap.plots.scatter(shap_values[:, "feature_name"])

# With interaction coloring
shap.plots.scatter(shap_values[:, "Age"], color=shap_values[:, "Income"])

Heatmap (Multiple Samples)

shap.plots.heatmap(shap_values[:100])

Common Patterns

Complete Analysis

import shap

# 1. Create explainer and compute
explainer = shap.TreeExplainer(model)
shap_values = explainer(X_test)

# 2. Global importance
shap.plots.beeswarm(shap_values)

# 3. Top feature relationships
shap.plots.scatter(shap_values[:, "top_feature"])

# 4. Individual explanation
shap.plots.waterfall(shap_values[0])

Compare Groups

# Compare feature importance across groups
group_a = X_test['category'] == 'A'
group_b = X_test['category'] == 'B'

shap.plots.bar({
    "Group A": shap_values[group_a],
    "Group B": shap_values[group_b]
})

Debug Errors

# Find misclassified samples
errors = model.predict(X_test) != y_test
error_idx = np.where(errors)[0]

# Explain why they failed
for idx in error_idx[:5]:
    shap.plots.waterfall(shap_values[idx])

Interpret Values

• Positive SHAP → Feature pushes prediction higher
• Negative SHAP → Feature pushes prediction lower
• Magnitude → Strength of impact
• Sum of SHAP values = Prediction - Baseline

Baseline: 0.30
Age: +0.15
Income: +0.10
Education: -0.05
Prediction: 0.30 + 0.15 + 0.10 - 0.05 = 0.50

Best Practices

1. Use TreeExplainer for tree models (fast, exact)
2. Use 100-1000 background samples for KernelExplainer
3. Start global (beeswarm) then go local (waterfall)
4. Check model output type (probability vs log-odds)
5. Validate with domain knowledge

vs Alternatives

Resources

• Docs: https://shap.readthedocs.io/
• Paper: Lundberg & Lee (2017) "A Unified Approach to Interpreting Model Predictions"