PFI（Permutation Feature Importance）による機械学習モデルの解釈

import numpy as np
import pandas as pd
from sklearn.linear_model import LinearRegression
from sklearn.metrics import mean_squared_error
from sklearn.preprocessing import StandardScaler
from sklearn.inspection import permutation_importance
from sklearn.model_selection import train_test_split
import matplotlib.pyplot as plt

# 1. データ生成
np.random.seed(0)
n_samples = 100
X = np.random.rand(n_samples, 3)
y = 0 * X[:, 0] + 1 * X[:, 1] + 2 * X[:, 2]  # f(x) = 0X_0 + 1X_1 + 2X_2

# 2. 学習データとテストデータに分割
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.3, random_state=0)

# 3. データの標準化（学習データに基づいて標準化）
scaler = StandardScaler()
X_train_std = scaler.fit_transform(X_train)
X_test_std = scaler.transform(X_test)

# 4. モデルの学習
model = LinearRegression()
model.fit(X_train_std, y_train)

# 5. テストデータを用いた PFI による特徴量重要度の計算
result = permutation_importance(model, X_test_std, y_test, n_repeats=30, random_state=0)
importance = result.importances_mean

# 6. 結果の可視化
feature_names = ['X_0', 'X_1', 'X_2']
importance_df = pd.DataFrame(importance, index=feature_names, columns=['Importance'])
importance_df.sort_values(by="Importance", ascending=True, inplace=True)

plt.figure(figsize=(8, 5))
plt.barh(importance_df.index, importance_df['Importance'], color='skyblue')
plt.xlabel("Feature Importance (PFI)")
plt.ylabel("Features")
plt.title("Permutation Feature Importance")
plt.show()

import numpy as np
import pandas as pd
from sklearn.linear_model import LinearRegression
from sklearn.metrics import mean_squared_error
from sklearn.preprocessing import StandardScaler
from sklearn.model_selection import train_test_split
import matplotlib.pyplot as plt

# 1. データ生成
np.random.seed(0)
n_samples = 100
X_0 = np.random.rand(n_samples)
X_1 = np.random.rand(n_samples)
X_2 = X_1 + np.random.normal(0, 0.01, n_samples)  # X_1とX_2を強く相関させる
X_3 = np.random.rand(n_samples)
X = np.column_stack((X_0, X_1, X_2, X_3))
y = 0 * X[:, 0] + 1 * X[:, 1] + 2 * X[:, 2] + 3 * X[:, 3]  # f(x) = 0X_0 + 1X_1 + 2X_2 + 3X_3

# 2. 学習データとテストデータに分割
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.3, random_state=0)

# 3. データの標準化（学習データに基づいて標準化）
scaler = StandardScaler()
X_train_std = scaler.fit_transform(X_train)
X_test_std = scaler.transform(X_test)

# 4. モデルの学習
model = LinearRegression()
model.fit(X_train_std, y_train)

# 5. GPFIの計算
# グループを定義 (X_1, X_2をグループ化)
groups = {
    "Group_1 (X_1, X_2)": [1, 2],  # インデックス1と2をグループ化
    "X_0": [0],
    "X_3": [3]
}

# ベースライン誤差の計算
baseline_mse = mean_squared_error(y_test, model.predict(X_test_std))

# グループごとにシャッフルして重要度を計算
importances = {}
for group_name, indices in groups.items():
    X_test_permuted = X_test_std.copy()
    for index in indices:
        np.random.shuffle(X_test_permuted[:, index])  # 特定のグループ内の特徴量をシャッフル
    permuted_mse = mean_squared_error(y_test, model.predict(X_test_permuted))
    importances[group_name] = permuted_mse - baseline_mse  # 誤差の増加分

# 6. 結果の可視化
importance_df = pd.DataFrame(list(importances.items()), columns=["Feature Group", "Importance"])
importance_df.sort_values(by="Importance", ascending=True, inplace=True)

plt.figure(figsize=(8, 5))
plt.barh(importance_df["Feature Group"], importance_df["Importance"], color="skyblue")
plt.xlabel("Feature Importance (GPFI)")
plt.ylabel("Feature Groups")
plt.title("Grouped Permutation Feature Importance")
plt.show()