加载自定义树模型到 SHAP 的示例

此笔记本展示了如何将自定义树集成模型传递到 SHAP 以进行解释。

import graphviz
import numpy as np
import scipy
import sklearn

import shap

简单回归树模型

这里我们定义了一个简单的回归树，然后将其作为自定义模型加载到 SHAP 中。

X, y = shap.datasets.adult()

orig_model = sklearn.tree.DecisionTreeRegressor(max_depth=2)
orig_model.fit(X, y)

DecisionTreeRegressor(max_depth=2)

在 Jupyter 环境中，请重新运行此单元格以显示 HTML 表示形式或信任该笔记本。
在 GitHub 上，HTML 表示形式无法渲染，请尝试使用 nbviewer.org 加载此页面。

dot_data = sklearn.tree.export_graphviz(orig_model, out_file=None, filled=True, rounded=True, special_characters=True)
graph = graphviz.Source(dot_data)
graph

有关这些属性的确切含义的更多信息，请参阅 scikit-learn 文档

# extract the arrays that define the tree
children_left = orig_model.tree_.children_left
children_right = orig_model.tree_.children_right
children_default = children_right.copy()  # because sklearn does not use missing values
features = orig_model.tree_.feature
thresholds = orig_model.tree_.threshold
values = orig_model.tree_.value.reshape(orig_model.tree_.value.shape[0], 1)
node_sample_weight = orig_model.tree_.weighted_n_node_samples

print("     children_left", children_left)  # note that negative children values mean this is a leaf node
print("    children_right", children_right)
print("  children_default", children_default)
print("          features", features)
print("        thresholds", thresholds.round(3))  # -2 means the node is a leaf node
print("            values", values.round(3))
print("node_sample_weight", node_sample_weight)

     children_left [ 1  2 -1 -1  5 -1 -1]
    children_right [ 4  3 -1 -1  6 -1 -1]
  children_default [ 4  3 -1 -1  6 -1 -1]
          features [ 5  8 -2 -2  2 -2 -2]
        thresholds [ 3.5000e+00  7.0735e+03 -2.0000e+00 -2.0000e+00  1.2500e+01 -2.0000e+00
 -2.0000e+00]
            values [[0.241]
 [0.066]
 [0.05 ]
 [0.962]
 [0.451]
 [0.335]
 [0.724]]
node_sample_weight [32561. 17800. 17482.   318. 14761. 10329.  4432.]

# define a custom tree model
tree_dict = {
    "children_left": children_left,
    "children_right": children_right,
    "children_default": children_default,
    "features": features,
    "thresholds": thresholds,
    "values": values,
    "node_sample_weight": node_sample_weight,
}
model = {"trees": [tree_dict]}

explainer = shap.TreeExplainer(model)

# Make sure that the ingested SHAP model (a TreeEnsemble object) makes the
# same predictions as the original model
assert np.abs(explainer.model.predict(X) - orig_model.predict(X)).max() < 1e-4

# make sure the SHAP values sum up to the model output (this is the local accuracy property)
assert np.abs(explainer.expected_value + explainer.shap_values(X).sum(1) - orig_model.predict(X)).max() < 1e-4

简单 GBM 分类模型（包含 2 棵树）

这里我们定义了一个简单的梯度提升分类器，然后将其作为自定义模型加载到 SHAP 中。

X2, y2 = shap.datasets.adult()
orig_model2 = sklearn.ensemble.GradientBoostingClassifier(n_estimators=2)
orig_model2.fit(X2, y2)

GradientBoostingClassifier(n_estimators=2)

在 Jupyter 环境中，请重新运行此单元格以显示 HTML 表示形式或信任该笔记本。
在 GitHub 上，HTML 表示形式无法渲染，请尝试使用 nbviewer.org 加载此页面。

拉取第一棵树的信息

tree_tmp = orig_model2.estimators_[0][0].tree_

# extract the arrays that define the tree
children_left1 = tree_tmp.children_left
children_right1 = tree_tmp.children_right
children_default1 = children_right1.copy()  # because sklearn does not use missing values
features1 = tree_tmp.feature
thresholds1 = tree_tmp.threshold
values1 = tree_tmp.value.reshape(tree_tmp.value.shape[0], 1)
node_sample_weight1 = tree_tmp.weighted_n_node_samples

print("     children_left1", children_left1)  # note that negative children values mean this is a leaf node
print("    children_right1", children_right1)
print("  children_default1", children_default1)
print("          features1", features1)
print("        thresholds1", thresholds1.round(3))
print("            values1", values1.round(3))
print("node_sample_weight1", node_sample_weight1)

     children_left1 [ 1  2  3 -1 -1  6 -1 -1  9 10 -1 -1 13 -1 -1]
    children_right1 [ 8  5  4 -1 -1  7 -1 -1 12 11 -1 -1 14 -1 -1]
  children_default1 [ 8  5  4 -1 -1  7 -1 -1 12 11 -1 -1 14 -1 -1]
          features1 [ 5  8  2 -2 -2  0 -2 -2  2  8 -2 -2  8 -2 -2]
        thresholds1 [ 3.5000e+00  7.0735e+03  1.2500e+01 -2.0000e+00 -2.0000e+00  2.0500e+01
 -2.0000e+00 -2.0000e+00  1.2500e+01  5.0955e+03 -2.0000e+00 -2.0000e+00
  5.0955e+03 -2.0000e+00 -2.0000e+00]
            values1 [[-0.   ]
 [-0.175]
 [-0.191]
 [-1.177]
 [-0.503]
 [ 0.721]
 [-0.223]
 [ 4.013]
 [ 0.211]
 [ 0.094]
 [ 0.325]
 [ 4.048]
 [ 0.483]
 [ 2.372]
 [ 4.128]]
node_sample_weight1 [3.2561e+04 1.7800e+04 1.7482e+04 1.4036e+04 3.4460e+03 3.1800e+02
 5.0000e+00 3.1300e+02 1.4761e+04 1.0329e+04 9.8070e+03 5.2200e+02
 4.4320e+03 3.7540e+03 6.7800e+02]

拉取第二棵树的信息

tree_tmp = orig_model2.estimators_[1][0].tree_

# extract the arrays that define the tree
children_left2 = tree_tmp.children_left
children_right2 = tree_tmp.children_right
children_default2 = children_right2.copy()  # because sklearn does not use missing values
features2 = tree_tmp.feature
thresholds2 = tree_tmp.threshold
values2 = tree_tmp.value.reshape(tree_tmp.value.shape[0], 1)
node_sample_weight2 = tree_tmp.weighted_n_node_samples

print("     children_left2", children_left2)  # note that negative children values mean this is a leaf node
print("    children_right2", children_right2)
print("  children_default2", children_default2)
print("          features2", features2)
print("        thresholds2", thresholds2.round(3))
print("            values2", values2.round(3))
print("node_sample_weight2", node_sample_weight2)

     children_left2 [ 1  2  3 -1 -1  6 -1 -1  9 10 -1 -1 13 -1 -1]
    children_right2 [ 8  5  4 -1 -1  7 -1 -1 12 11 -1 -1 14 -1 -1]
  children_default2 [ 8  5  4 -1 -1  7 -1 -1 12 11 -1 -1 14 -1 -1]
          features2 [ 5  8  2 -2 -2  0 -2 -2  2  8 -2 -2  8 -2 -2]
        thresholds2 [ 3.5000e+00  7.0735e+03  1.3500e+01 -2.0000e+00 -2.0000e+00  2.0500e+01
 -2.0000e+00 -2.0000e+00  1.2500e+01  5.0955e+03 -2.0000e+00 -2.0000e+00
  5.0955e+03 -2.0000e+00 -2.0000e+00]
            values2 [[-1.000e-03]
 [-1.580e-01]
 [-1.720e-01]
 [-1.062e+00]
 [ 1.360e-01]
 [ 6.420e-01]
 [-2.030e-01]
 [ 2.993e+00]
 [ 1.880e-01]
 [ 8.400e-02]
 [ 2.870e-01]
 [ 3.015e+00]
 [ 4.310e-01]
 [ 1.895e+00]
 [ 3.066e+00]]
node_sample_weight2 [3.2561e+04 1.7800e+04 1.7482e+04 1.6560e+04 9.2200e+02 3.1800e+02
 5.0000e+00 3.1300e+02 1.4761e+04 1.0329e+04 9.8070e+03 5.2200e+02
 4.4320e+03 3.7540e+03 6.7800e+02]

创建 SHAP 树列表

# define a custom tree model
tree_dicts = [
    {
        "children_left": children_left1,
        "children_right": children_right1,
        "children_default": children_default1,
        "features": features1,
        "thresholds": thresholds1,
        "values": values1 * orig_model2.learning_rate,
        "node_sample_weight": node_sample_weight1,
    },
    {
        "children_left": children_left2,
        "children_right": children_right2,
        "children_default": children_default2,
        "features": features2,
        "thresholds": thresholds2,
        "values": values2 * orig_model2.learning_rate,
        "node_sample_weight": node_sample_weight2,
    },
]
model2 = {
    "trees": tree_dicts,
    "base_offset": scipy.special.logit(orig_model2.init_.class_prior_[1]),
    "tree_output": "log_odds",
    "objective": "binary_crossentropy",
    "input_dtype": np.float32,  # this is what type the model uses the input feature data
    "internal_dtype": np.float64,  # this is what type the model uses for values and thresholds
}

解释自定义模型

# build a background dataset for us to use based on people near a 0.95 cutoff
vs = np.abs(orig_model2.predict_proba(X2)[:, 1] - 0.95)
inds = np.argsort(vs)
inds = inds[:200]

# build an explainer that explains the probability output of the model
explainer2 = shap.TreeExplainer(
    model2,
    X2.iloc[inds, :],
    feature_perturbation="interventional",
    model_output="probability",
)

# Make sure that the ingested SHAP model (a TreeEnsemble object) makes the
# same predictions as the original model
assert np.abs(explainer2.model.predict(X2, output="probability") - orig_model2.predict_proba(X2)[:, 1]).max() < 1e-4

# make sure the sum of the SHAP values equals the model output
shap_sum = explainer2.expected_value + explainer2.shap_values(X2.iloc[:, :]).sum(1)
assert np.abs(shap_sum - orig_model2.predict_proba(X2)[:, 1]).max() < 1e-4