@浙大疏锦行
方差筛选是最基础的过滤法：计算特征的方差，剔除方差极低的特征（这类特征数值变化小，对样本区分度弱）。优点是计算极快，缺点是只看特征自身，不考虑和目标的关联。
皮尔逊相关系数筛选属于过滤法：计算特征与目标变量的皮尔逊相关系数（衡量线性相关程度），保留相关系数绝对值高的特征。优点是直观易懂，缺点仅能捕捉线性关联，对非线性关系无效。
Lasso 筛选属于嵌入法：在线性模型（如线性回归、逻辑回归）中加入 L1 正则化，迫使不重要的特征系数收缩至 0，最终保留系数非 0 的特征。既能完成特征筛选，也能同步训练模型，适合高维数据。
树模型重要性属于嵌入法：基于决策树 / 随机森林 / XGBoost 等树模型，通过 “特征对节点分裂的贡献度” 计算重要性，保留重要性高的特征。优点是能捕捉非线性关联，缺点是易受高基数特征干扰。
SHAP 重要性是树模型重要性的进阶版：基于 SHAP 值（解释模型预测的统一框架）计算特征重要性，不仅能体现特征的整体影响程度，还能展示特征对预测结果的正负方向，解释性更强。
递归特征消除（RFE）属于包裹法：反复训练指定模型，每次移除模型判定的 “最不重要特征”，直到保留预设数量的特征。优点是直接以模型性能为筛选依据，精度较高；缺点是计算成本高（需多次训练模型）
作业：对心脏病数据集完成特征筛选，并对比精度。import pandas as pd import numpy as np import matplotlib.pyplot as plt import seaborn as sns from sklearn.model_selection import train_test_split from sklearn.preprocessing import StandardScaler from sklearn.linear_model import LogisticRegression from sklearn.ensemble import RandomForestClassifier from sklearn.feature_selection import VarianceThreshold, RFE from sklearn.metrics import accuracy_score, classification_report import shap import xgboost as xgb # ---------------------- 1. 加载数据与预处理 ---------------------- # 加载数据集 df = pd.read_csv('E:/PythonStudy/data.csv') # 分离特征与目标变量 X = df.drop('target', axis=1) y = df['target'] # 填充缺失值 X = X.fillna(X.mean()) # ---------------------- 2. 数据探索可视化 ---------------------- def data_explore(df): """数据分布与特征相关性可视化""" # 目标变量分布饼图 target_count = df["target"].value_counts().reset_index() target_count["target"] = target_count["target"].map(lambda x: "患病" if x == 1 else "正常") target_count["占比"] = target_count["count"] / target_count["count"].sum() * 100 plt.figure(figsize=(6, 6)) wedges, texts = plt.pie( target_count["count"], wedgeprops={"width": 0.4, "edgecolor": "#000"}, labels=target_count["target"], autopct="%1.1f%%" ) plt.title("样本患病情况分布", fontsize=15, pad=20) plt.show() # 特征与目标变量的相关性热力图 corr = df.corr() plt.figure(figsize=(12, 8)) sns.heatmap( corr, annot=True, cmap="RdBu_r", vmin=-1, vmax=1, annot_kws={"fontsize": 10} ) plt.title("特征-目标相关性热力图", fontsize=15) plt.show() # 执行数据探索 data_explore(df) # ---------------------- 3. 数据集划分与标准化 ---------------------- X_train, X_test, y_train, y_test = train_test_split( X, y, test_size=0.2, random_state=42, stratify=y ) scaler = StandardScaler() X_train_scaled = pd.DataFrame(scaler.fit_transform(X_train), columns=X.columns) X_test_scaled = pd.DataFrame(scaler.transform(X_test), columns=X.columns) # ---------------------- 4. 特征筛选方法定义 ---------------------- def select_by_variance(X_train, X_test, threshold=0.1): """方差筛选：移除低方差特征""" vt = VarianceThreshold(threshold) X_train_sel = vt.fit_transform(X_train) X_test_sel = vt.transform(X_test) cols = X_train.columns[vt.get_support()] return pd.DataFrame(X_train_sel, columns=cols), pd.DataFrame(X_test_sel, columns=cols) def select_by_pearson(X_train, y_train, X_test, top_n=5): """皮尔逊相关系数筛选：保留与目标相关性topN的特征""" corr = [abs(X_train[col].corr(y_train)) for col in X_train.columns] top_cols = X_train.columns[np.argsort(corr)[-top_n:]] return X_train[top_cols], X_test[top_cols] def select_by_lasso(X_train_scaled, y_train, X_test_scaled): """Lasso筛选：保留系数非0的特征""" lasso = LogisticRegression(penalty='l1', solver='liblinear', C=0.1, random_state=42) lasso.fit(X_train_scaled, y_train) top_cols = X_train_scaled.columns[lasso.coef_[0] != 0] return X_train_scaled[top_cols], X_test_scaled[top_cols] def select_by_tree_importance(X_train, y_train, X_test, top_n=5): """树模型重要性筛选：保留随机森林重要性topN的特征""" rf = RandomForestClassifier(random_state=42) rf.fit(X_train, y_train) top_cols = X_train.columns[np.argsort(rf.feature_importances_)[-top_n:]] return X_train[top_cols], X_test[top_cols] def select_by_shap(X_train, y_train, X_test, top_n=5): """SHAP重要性筛选：保留SHAP值topN的特征""" rf = RandomForestClassifier(random_state=42) rf.fit(X_train, y_train) explainer = shap.TreeExplainer(rf) shap_values = explainer.shap_values(X_train) shap_importance = np.mean(np.abs(shap_values[1]), axis=0) top_cols = X_train.columns[np.argsort(shap_importance)[-top_n:]] return X_train[top_cols], X_test[top_cols] def select_by_rfe(X_train, y_train, X_test, n_features=5): """递归特征消除：保留指定数量的特征""" lr = LogisticRegression(max_iter=1000, random_state=42) rfe = RFE(estimator=lr, n_features_to_select=n_features) rfe.fit(X_train, y_train) top_cols = X_train.columns[rfe.get_support()] return X_train[top_cols], X_test[top_cols] # ---------------------- 5. 模型评估与结果记录 ---------------------- # 待对比的模型 models = { "逻辑回归": LogisticRegression(max_iter=1000, random_state=42), "XGBoost": xgb.XGBClassifier(n_estimators=100, random_state=42) } # 待对比的特征筛选方法 methods = { "方差筛选": select_by_variance, "皮尔逊相关": select_by_pearson, "Lasso": select_by_lasso, "树模型重要性": select_by_tree_importance, "SHAP重要性": select_by_shap, "RFE": select_by_rfe } # 存储所有评估结果 all_results = {} # 遍历筛选方法与模型，计算指标 for method_name, method_func in methods.items(): # 执行特征筛选 if method_name == "Lasso": X_train_sel, X_test_sel = method_func(X_train_scaled, y_train, X_test_scaled) else: X_train_sel, X_test_sel = method_func(X_train, y_train, X_test) # 评估不同模型 method_res = {} for model_name, model in models.items(): model.fit(X_train_sel, y_train) y_pred = model.predict(X_test_sel) # 计算准确率与F1值 acc = accuracy_score(y_test, y_pred) f1 = classification_report(y_test, y_pred, output_dict=True)["weighted avg"]["f1-score"] method_res[model_name] = {"准确率": round(acc, 4), "F1值": round(f1, 4)} print(f"{method_name}-{model_name}：准确率={acc:.4f}，F1={f1:.4f}") all_results[method_name] = method_res print("-" * 60) # ---------------------- 6. 结果可视化 ---------------------- def plot_accuracy_compare(all_results): """绘制各筛选方法下的模型准确率对比图""" plt.figure(figsize=(12, 6)) x = np.arange(len(all_results)) width = 0.35 # 提取不同模型的准确率 lr_acc = [all_results[method]["逻辑回归"]["准确率"] for method in all_results] xgb_acc = [all_results[method]["XGBoost"]["准确率"] for method in all_results] # 绘制柱状图 plt.bar(x - width/2, lr_acc, width, label="逻辑回归", color="#1f77b4") plt.bar(x + width/2, xgb_acc, width, label="XGBoost", color="#ff7f0e") # 图表配置 plt.xticks(x, all_results.keys(), rotation=45) plt.ylabel("准确率", fontsize=12) plt.title("不同特征筛选方法的模型准确率对比", fontsize=15) plt.legend() plt.tight_layout() plt.show() # 生成可视化结果 plot_accuracy_compare(all_results)
