Explaining Model Predictions with SHAP

What a SHAP value represents#

SHAP applies the Shapley-value idea from cooperative game theory to feature attribution. For an individual prediction, it assigns each feature an additive contribution relative to a reference value, averaging the feature’s marginal contribution across possible feature subsets under the chosen explanation setup.

This provides a consistent accounting framework for a model output, but the result depends on the model, background distribution, and assumptions used to handle missing or dependent features. An attribution explains how a model’s prediction changes under that setup; it does not show that changing the feature would cause the real-world outcome to change.

Explainers and plots in the tutorial#

The source introduces several explainers:

• TreeExplainer for tree-based models.
• LinearExplainer for linear models.
• KernelExplainer as a model-agnostic option, with higher computational cost.
• DeepExplainer for supported deep-learning models.

It also walks through visualizations such as force plots and summary-style plots to inspect individual and global patterns. The appropriate explainer and plot depend on the model and the question; global summaries can hide variation between cases.

Read explanations with care#

Feature correlation, reference-data choices, model limitations, and implementation details can all affect interpretation. SHAP is one tool for auditing model behavior, not a substitute for validation, domain review, or causal analysis.

See the Chinese article for formulas, code examples, and additional visualization walkthroughs.