nexus/wiki/concepts/SHAP-Analysis.md

title, type, tags, sources, last_updated

title	type	tags	sources	last_updated
SHAP Analysis	concept	interpretability explainability ml-ops	specialized-model-qa	2026-04-20

Definition

SHAP（SHapley Additive exPlanations）是一种基于博弈论的模型可解释性方法，用于分解单个预测中每个特征的贡献。

Outputs

• 全局解释：summary / beeswarm / bar plot
• 局部解释：waterfall / force plot
• 交互分析：SHAP interaction values

Use in Model QA

• 解释预测驱动因素
• 检查特征贡献是否稳定
• 识别异常依赖或潜在偏差

Related Concepts

• Partial Dependence Plots
• Model Audit