如何输出概率中的 Shap 值并从二元分类器制作force_plot

Question

我需要绘制每个特征如何影响我的 LightGBM 二元分类器中每个样本的预测概率。所以我需要以概率的形式输出Shap值，而不是正常的Shap值。它似乎没有任何概率输出选项。

下面的示例代码是我用来生成 Shap 值的数据帧并为第一个数据样本执行 force_plot 的代码。有谁知道我应该如何修改代码来改变输出？ 我是 Shap 值和 Shap 包的新手。预先非常感谢。

import pandas as pd
import numpy as np
import shap
import lightgbm as lgbm
from sklearn.model_selection import train_test_split
from sklearn.datasets import load_breast_cancer

data = load_breast_cancer()
X = pd.DataFrame(data.data, columns=data.feature_names)
y = data.target
X_train, X_test, y_train, y_test = train_test_split(X, y,  test_size=0.2)
model = lgbm.LGBMClassifier()
model.fit(X_train, y_train)


explainer = shap.TreeExplainer(model)
shap_values = explainer(X_train)

# force plot of first row for class 1
class_idx = 1
row_idx = 0
expected_value = explainer.expected_value[class_idx]
shap_value = shap_values[:,:,class_idx].values[row_idx]

shap.force_plot (base_value = expected_value,  shap_values = shap_value, features = X_train.iloc[row_idx, :], matplotlib=True)

# dataframe of shap values for class 1
shap_df = pd.DataFrame(shap_values[:,:, 1 ].values, columns = shap_values.feature_names)

Answer 1

TL;DR：

您可以使用 force_plot 方法中的 link="logit" 在概率空间中获得绘图结果：

import pandas as pd
import numpy as np
import shap
import lightgbm as lgbm
from sklearn.model_selection import train_test_split
from sklearn.datasets import load_breast_cancer
from scipy.special import expit

shap.initjs()

data = load_breast_cancer()

X = pd.DataFrame(data.data, columns=data.feature_names)
y = data.target
X_train, X_test, y_train, y_test = train_test_split(
    X, y, test_size=0.2, random_state=42
)

model = lgbm.LGBMClassifier()
model.fit(X_train, y_train)

explainer_raw = shap.TreeExplainer(model)
shap_values = explainer_raw(X_train)

# force plot of first row for class 1
class_idx = 1
row_idx = 0
expected_value = explainer_raw.expected_value[class_idx]
shap_value = shap_values[:, :, class_idx].values[row_idx]

shap.force_plot(
    base_value=expected_value,
    shap_values=shap_value,
    features=X_train.iloc[row_idx, :],
    link="logit",
)

预期输出：

或者，您可以通过以下明确实现相同的效果指定您有兴趣解释的 model_output="probability"：

explainer = shap.TreeExplainer(
    model,
    data=X_train,
    feature_perturbation="interventional",
    model_output="probability",
)
shap_values = explainer(X_train)

# force plot of first row for class 1
class_idx = 1
row_idx = 0

shap_value = shap_values.values[row_idx]

shap.force_plot(
    base_value=expected_value, 
    shap_values=shap_value, 
    features=X_train.iloc[row_idx, :]
)

预期输出：

< img src="https://i.sstatic.net/YJRYv.png" alt="在此处输入图像描述">

但是，了解此处发生的情况以找出这些数字的来源可能更有趣来自：

1. 我们的目标概率兴趣点：

model_proba= model.predict_proba(X_train.iloc[[row_idx]])
model_proba
# array([[0.00275887, 0.99724113]])

2. 来自给定 X_train 作为背景的模型的基本案例原始数据（注意，LightGBM 输出类 1 的原始数据）：

model.predict(X_train, raw_score=True).mean()
# 2.4839751932445577

3. 来自 X_train 的原始基本案例code>SHAP （注意，它们是对称的）：

bv = explainer_raw(X_train).base_values[0]
bv
# array([-2.48397519,  2.48397519])

4. 感兴趣点的原始 SHAP 值：

sv_0 = explainer_raw(X_train).values[row_idx].sum(0)
sv_0
# array([-3.40619584,  3.40619584])

5. 从 SHAP 值推断的 Proba（通过 sigmoid）：

shap_proba = expit(bv + sv_0)
shap_proba
# array([0.00275887, 0.99724113])

6. 检查：

assert np.allclose(model_proba, shap_proba)

如果有不清楚的地方，请提问。

旁注

如果您正在分析不同特征的原始尺寸效应，Proba 可能会产生误导，因为 sigmoid 是非线性的，并且在达到特定阈值后会饱和。

有些人希望在概率空间中看到 SHAP 值，但这是不可行的，因为：

• SHAP 值是通过构造进行加法的（准确地说，Shapley 加法解释是对所有可能的特征联盟的平均边际贡献）
• exp(a + b) != exp(a) + exp(b)

您可能会发现有用：

1. 二元分类中的特征重要性并仅提取其中一个类的 SHAP 值答案< /p>

2. 使用SHAP时如何解释GBT分类器的base_value？ 答案< /p>

Answer 2

您可以考虑通过 softmax() 函数运行输出值。作为参考，它被定义为：

def get_softmax_probabilities(x):
    return np.exp(x) / np.sum(np.exp(x)).reshape(-1, 1)

并且还有一个 scipy 实现：

from scipy.special import softmax

softmax() 的输出将是与向量 x 中的（相对）值成比例的概率，即您的商店值。

Answer 3

import pandas as pd
import numpy as np
import shap
import lightgbm as lgbm
from sklearn.model_selection import train_test_split
from sklearn.datasets import load_breast_cancer

data = load_breast_cancer()
X = pd.DataFrame(data.data, columns=data.feature_names)
y = data.target
X_train, X_test, y_train, y_test = train_test_split(X, y,  test_size=0.2)
print('X_train: ',X_train.shape)
print('X_test: ',X_test.shape)

model = lgbm.LGBMClassifier()
model.fit(X_train, y_train)

explainer = shap.TreeExplainer(model)
shap_values = explainer.shap_values(X_train)

# plot
# shap.summary_plot(shap_values[class_idx], X_train, plot_type='bar')
# shap.summary_plot(shap_values[class_idx], X_train)

# shap_value = shap_values[:,:,class_idx].values[row_idx]
# shap.force_plot (base_value = expected_value,  shap_values = shap_value, features = X_train.iloc[row_idx, :], matplotlib=True)
# # dataframe of shap values for class 1
# shap_df = pd.DataFrame(shap_values[:,:, 1 ].values, columns = shap_values.feature_names)

# verification
def verification(index_number,class_idx):
    print('-----------------------------------')
    print('index_number: ', index_number)
    print('class_idx: ', class_idx)
    print('')
    
    y_base = explainer.expected_value[class_idx]
    print('y_base: ', y_base)

    player_explainer = pd.DataFrame()
    player_explainer['feature_value'] = X_train.iloc[j].values
    player_explainer['shap_value'] = shap_values[class_idx][j]
    print('verification: ')
    print('y_base + sum_of_shap_values: %.2f'%(y_base + player_explainer['shap_value'].sum()))
    print('y_pred: %.2f'%(y_train[j]))

j = 10  # index
verification(j,0)
verification(j,1)

# show: 
# X_train:  (455, 30)
# X_test:  (114, 30)
# -----------------------------------
# index_number:  10
# class_idx:  0

# y_base:  -2.391423081639827
# verification: 
# y_base + sum_of_shap_values: -9.40
# y_pred: 1.00
# -----------------------------------
# index_number:  10
# class_idx:  1

# y_base:  2.391423081639827
# verification: 
# y_base + sum_of_shap_values: 9.40
# y_pred: 1.00
# -9.40,9.40 takes the maximum value（class_idx:1 = y_pred）, and the result is obviously correct.

我帮助您实现了这一目标并验证了结果的可靠性。