使用自定义掩码器

这里我们展示了如何为任何 SHAP 模型不可知解释方法提供自定义的 Masker。掩码通常是领域相关的，因此考虑超越 SHAP 默认包含的扰动数据的其他方法通常是有帮助的。

import xgboost

import shap

# train XGBoost model
X, y = shap.datasets.adult()
model = xgboost.XGBClassifier().fit(X.values, y)


# A masking function takes a binary mask vector as the first argument and
# the model arguments for a single sample after that
# It returns a masked version of the input x, where you can return multiple
# rows to average over a distribution of masking types
def custom_masker(mask, x):
    # in this simple example we just zero out the features we are masking
    return (x * mask).reshape(1, len(x))


# compute SHAP values
explainer = shap.Explainer(model.predict_proba, custom_masker)
shap_values = explainer(X[:100])

# plot the SHAP values for the positive class
shap.plots.beeswarm(shap_values[..., 1])

---

有更多有用示例的想法吗？我们鼓励提交增加此文档笔记本的拉取请求！