本文演示了如何结合 KMeans 聚类、随机森林分类器和 SHAP 值，对心脏病数据集进行特征重要性分析。核心流程包括数据预处理、聚类评估、特征重要性排序和可视化。

1. 数据准备

import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns
import warnings
from sklearn.model_selection import train_test_split

warnings.filterwarnings("ignore")
# 设置中文字体
plt.rcParams['font.sans-serif'] = ['SimHei']
plt.rcParams['axes.unicode_minus'] = False

# 加载数据
df = pd.read_csv('heart.csv')

# 处理缺失值：使用中位数填充
missing_features = df.columns[df.isnull().any()].tolist()
for feat in missing_features:
    median_val = df[feat].median()
    df[feat] = df[feat].fillna(median_val)

# 分割特征和标签
features = df.drop('target', axis=1)
target = df['target']

# 划分训练集和测试集 (8:2)
X_train, X_test, y_train, y_test = train_test_split(features, target, test_size=0.2, random_state=42)

2. 数据标准化与聚类评估

from sklearn.cluster import KMeans
from sklearn.preprocessing import StandardScaler
from sklearn.decomposition import PCA
from sklearn.metrics import silhouette_score, calinski_harabasz_score, davies_bouldin_score

# 标准化
scaler = StandardScaler()
scaled_data = scaler.fit_transform(features)

# 评估不同 k 值 (2-10)
k_values = range(2, 11)
inertia_list = []
sil_list = []
ch_list = []
db_list = []

for k in k_values:
    km = KMeans(n_clusters=k, random_state=42)
    labels = km.fit_predict(scaled_data)
    inertia_list.append(km.inertia_)
    sil_list.append(silhouette_score(scaled_data, labels))
    ch_list.append(calinski_harabasz_score(scaled_data, labels))
    db_list.append(davies_bouldin_score(scaled_data, labels))
    print(f"k={k}, 惯性:{km.inertia_:.2f}, 轮廓:{silhouette_score(scaled_data, labels):.3f}, CH:{ch_list[-1]:.2f}, DB:{db_list[-1]:.3f}")

# 选择 k=3（实际应用中可根据肘部法则或指标选择）
final_k = 3
kmeans = KMeans(n_clusters=final_k, random_state=42)
cluster_labels = kmeans.fit_predict(scaled_data)
features['Cluster'] = cluster_labels

# PCA 降维可视化
pca = PCA(n_components=2)
pca_data = pca.fit_transform(scaled_data)

plt.figure(figsize=(6, 5))
sns.scatterplot(x=pca_data[:, 0], y=pca_data[:, 1], hue=cluster_labels, palette='viridis')
plt.title(f'KMeans 聚类结果 (k={final_k})')
plt.xlabel('主成分1')
plt.ylabel('主成分2')
plt.show()

print("各簇样本数：")
print(features['Cluster'].value_counts())

3. 使用随机森林和 SHAP 分析特征重要性

首先构建随机森林模型，然后用 SHAP 解释器计算特征重要性。

import shap
from sklearn.ensemble import RandomForestClassifier

# 除去聚类标签列
X_cluster = features.drop('Cluster', axis=1)
y_cluster = features['Cluster']

# 训练随机森林
rf_model = RandomForestClassifier(n_estimators=100, random_state=42)
rf_model.fit(X_cluster, y_cluster)

# SHAP 解释器
explainer = shap.TreeExplainer(rf_model)
shap_vals = explainer.shap_values(X_cluster)

# 处理多维 SHAP 值（多分类）
shap_vals_array = np.array(shap_vals)
print(f"SHAP 值形状: {shap_vals_array.shape}")

# 绘制特征重要性条形图（取第一个类别）
shap.summary_plot(shap_vals[0], X_cluster, plot_type="bar", show=False)
plt.title("SHAP 特征重要性（条形图）")
plt.show()

4. 选定特征分布分析

选取年龄、性别、胸痛类型、静息血压四个特征进行分布分析。

selected_feats = ['age', 'sex', 'cp', 'trestbps']

# 判断变量类型
for feat in selected_feats:
    unique_cnt = df[feat].nunique()
    feat_type = "离散型" if unique_cnt < 10 else "连续型"
    print(f"{feat}: {unique_cnt} 个唯一值 -> {feat_type}")

# 总样本分布
fig, axes = plt.subplots(2, 2, figsize=(12, 8))
for idx, feat in enumerate(selected_feats):
    ax = axes[idx//2, idx%2]
    ax.hist(df[feat], bins=20)
    ax.set_title(f'{feat} 分布')
    ax.set_xlabel(feat)
    ax.set_ylabel('频数')
plt.tight_layout()
plt.show()

5. 按簇分析特征分布

# 按簇拆分数
clusters = features['Cluster'].unique()
for cluster in clusters:
    subset = features[features['Cluster'] == cluster]
    fig, axes = plt.subplots(2, 2, figsize=(12, 8))
    for idx, feat in enumerate(selected_feats):
        ax = axes[idx//2, idx%2]
        ax.hist(subset[feat], bins=20)
        ax.set_title(f'簇 {cluster} - {feat} 分布')
        ax.set_xlabel(feat)
        ax.set_ylabel('频数')
    plt.tight_layout()
    plt.show()

通过以上步骤，我们成功将数据分为 3 个簇，并利用 SHAP 和随机森林识别出对聚类贡献最大的特征，进一步观察了这些特征在不同簇中的分布差异。这有助于理解数据的内部结构。