Python实操之RBF-SVM分类:把标准化、调参和测试集评估放对位置

import osfrom pathlib import Path# 把Matplotlib缓存放到可写目录，保证后台环境绘图稳定os.environ[”MPLCONFIGDIR”] = ”/tmp/wechat_python_svm_mpl_20260811”import matplotlibmatplotlib.use(”Agg”)import matplotlib.pyplot as pltimport numpy as npimport pandas as pdimport sklearnfrom sklearn.datasets import load_breast_cancerfrom sklearn.metrics import balanced_accuracy_score, confusion_matrix, recall_scorefrom sklearn.model_selection import GridSearchCV, StratifiedKFold, train_test_splitfrom sklearn.pipeline import Pipelinefrom sklearn.preprocessing import StandardScalerfrom sklearn.svm import SVC# 固定数据划分与交叉验证随机种子，保证结果可复现RANDOM_SEED = 20260811dataset = load_breast_cancer(as_frame=True)X = dataset.datay = dataset.target# 先留出测试集，后续调参只能接触训练集X_train, X_test, y_train, y_test = train_test_split(    X,    y,    test_size=0.25,    stratify=y,    random_state=RANDOM_SEED,)# 类别0是malignant，类别1是benign，输出分布便于解释指标train_counts = y_train.value_counts().sort_index()test_counts = y_test.value_counts().sort_index()print(”数据来源：scikit-learn内置Breast Cancer Wisconsin数据”)print(f”样本数 / 特征数：{X.shape[0]} / {X.shape[1]}”)print(f”完整类别分布：malignant={(y == 0).sum()}，benign={(y == 1).sum()}”)print(f”训练集：{len(X_train)}；测试集：{len(X_test)}”)print(f”scikit-learn版本：{sklearn.__version__}”)

数据来源：scikit-learn内置Breast Cancer Wisconsin数据样本数 / 特征数：569 / 30完整类别分布：malignant=212，benign=357训练集：426；测试集：143scikit-learn版本：1.9.0

# 标准化放进Pipeline，确保每个交叉验证折只使用自己的训练部分拟合尺度svm_pipeline = Pipeline(    [        (”scaler”, StandardScaler()),        (”svc”, SVC(kernel=”rbf”, class_weight=”balanced”)),    ])# C控制正则化权衡，gamma控制单个样本影响范围c_values = [0.1, 1.0, 10.0, 100.0]gamma_values = [0.0001, 0.001, 0.01, 0.1]param_grid = {    ”svc__C”: c_values,    ”svc__gamma”: gamma_values,}# 调参只在训练集内做五折分层交叉验证，评分使用平衡准确率cv = StratifiedKFold(n_splits=5, shuffle=True, random_state=RANDOM_SEED)grid_search = GridSearchCV(    estimator=svm_pipeline,    param_grid=param_grid,    scoring=”balanced_accuracy”,    cv=cv,    n_jobs=1,    refit=True,    return_train_score=True,)grid_search.fit(X_train, y_train)print(f”候选组合：{len(c_values) * len(gamma_values)} 个”)print(f”交叉验证：{cv.get_n_splits()} 折分层划分”)print(f”最佳C：{grid_search.best_params_['svc__C']}”)print(f”最佳gamma：{grid_search.best_params_['svc__gamma']}”)print(f”最佳CV平衡准确率：{grid_search.best_score_:.3f}”)

候选组合：16 个交叉验证：5 折分层划分最佳C：10.0最佳gamma：0.001最佳CV平衡准确率：0.969

# 将每组参数的验证分数整理为C乘gamma矩阵cv_results = pd.DataFrame(grid_search.cv_results_)score_matrix = np.zeros((len(gamma_values), len(c_values)))for row_id, gamma in enumerate(gamma_values):    for col_id, c_value in enumerate(c_values):        matched = cv_results[            (cv_results[”param_svc__gamma”] == gamma)            & (cv_results[”param_svc__C”] == c_value)        ]        score_matrix[row_id, col_id] = matched[”mean_test_score”].iloc[0]# 热图展示参数区域，而不只报告单个最佳点fig, ax = plt.subplots(figsize=(8.2, 5.6))image = ax.imshow(score_matrix, cmap=”YlGnBu”, vmin=score_matrix.min() - 0.005, vmax=score_matrix.max())ax.set_title(”RBF-SVM Cross-validated Balanced Accuracy”, pad=14, weight=”bold”)ax.set_xlabel(”C”)ax.set_ylabel(”Gamma”)ax.set_xticks(range(len(c_values)), [str(value) for value in c_values])ax.set_yticks(range(len(gamma_values)), [str(value) for value in gamma_values])for row_id in range(score_matrix.shape[0]):    for col_id in range(score_matrix.shape[1]):        ax.text(col_id, row_id, f”{score_matrix[row_id, col_id]:.3f}”, ha=”center”, va=”center”, weight=”bold”)fig.colorbar(image, ax=ax, label=”Mean CV balanced accuracy”)fig.tight_layout()grid_plot = Path(”/tmp/python_svm_grid_heatmap.png”)fig.savefig(grid_plot, dpi=180, bbox_inches=”tight”)plt.close(fig)# 训练分数与验证分数的差距只作为过拟合线索，不替代测试集评估best_index = grid_search.best_index_best_train_score = cv_results.loc[best_index, ”mean_train_score”]cv_gap = best_train_score - grid_search.best_score_print(f”参数网格最高分：{score_matrix.max():.3f}”)print(f”参数网格最低分：{score_matrix.min():.3f}”)print(f”最佳组合训练-CV差距：{cv_gap:.3f}”)print(f”Grid chart：{grid_plot.name} ({grid_plot.stat().st_size:,} bytes)”)

参数网格最高分：0.969参数网格最低分：0.500最佳组合训练-CV差距：0.012Grid chart：python_svm_grid_heatmap.png (69,131 bytes)