第 8 章 · 乳腺癌 MLP 分类代码实验¶

# 加载 Wisconsin 乳腺癌数据：这是二分类模型的经典公开案例。
cancer = load_breast_cancer(as_frame=True)
cancer_label_names = np.array(["恶性", "良性"])
X_mlp = cancer.data
y_mlp = cancer.target

X_train_mlp, X_test_mlp, y_train_mlp, y_test_mlp = train_test_split(
    X_mlp,
    y_mlp,
    test_size=0.25,
    stratify=y_mlp,
    random_state=8,
)

feature_preview = X_mlp[[
    "mean radius",
    "mean texture",
    "mean perimeter",
    "mean area",
    "mean smoothness",
]].head(8)

cancer_summary = pd.DataFrame(
    {
        "样本数": [len(X_mlp)],
        "输入维度": [X_mlp.shape[1]],
        "类别数": [len(np.unique(y_mlp))],
        "训练样本": [len(X_train_mlp)],
        "测试样本": [len(X_test_mlp)],
        "类别名称": [", ".join(cancer.target_names)],
    }
)
display(cancer_summary)
display(feature_preview.round(3))

# 训练小型 MLP，并读取每轮损失。
mlp = make_pipeline(
    StandardScaler(),
    MLPClassifier(
        hidden_layer_sizes=(32,),
        activation="relu",
        solver="adam",
        learning_rate_init=0.01,
        max_iter=80,
        random_state=8,
    ),
)

mlp.fit(X_train_mlp, y_train_mlp)
classifier = mlp.named_steps["mlpclassifier"]
mlp_trace = pd.DataFrame(
    {
        "轮次": np.arange(1, len(classifier.loss_curve_) + 1),
        "损失": classifier.loss_curve_,
    }
).round(4)

display(mlp_trace.iloc[[0, 1, 2, 9, 19, len(mlp_trace) - 1]])

# 查看测试集预测、准确率和混淆矩阵。
test_pred_mlp = mlp.predict(X_test_mlp)
test_prob_mlp = mlp.predict_proba(X_test_mlp)

score_df = pd.DataFrame(
    [{"数据": "乳腺癌测试集", "准确率": accuracy_score(y_test_mlp, test_pred_mlp)}]
).round(3)
confusion_df = pd.DataFrame(
    confusion_matrix(y_test_mlp, test_pred_mlp),
    index=[f"真实_{name}" for name in cancer_label_names],
    columns=[f"预测_{name}" for name in cancer_label_names],
)
sample_result = pd.DataFrame(
    {
        "样本编号": X_test_mlp.index[:12],
        "真实": y_test_mlp.to_numpy()[:12],
        "预测": test_pred_mlp[:12],
        "预测置信度": np.max(test_prob_mlp[:12], axis=1),
    }
).round(3)

display(score_df)
display(sample_result)
display(confusion_df)

# 绘制损失曲线、混淆矩阵和预测置信度。
fig, axes = plt.subplots(1, 3, figsize=(12.0, 4.4), gridspec_kw={"width_ratios": [1.05, 1.0, 1.15]})

axes[0].plot(mlp_trace["轮次"], mlp_trace["损失"], color="#2563eb", linewidth=2.4)
axes[0].set_title("MLP 训练损失", loc="left", fontweight="bold")
axes[0].set_xlabel("轮次")
axes[0].set_ylabel("损失")
axes[0].grid(True, color="#e2e8f0", linewidth=0.8)

im = axes[1].imshow(confusion_df.to_numpy(), cmap="YlGnBu")
axes[1].set_title("测试集混淆矩阵", loc="left", fontweight="bold")
axes[1].set_xticks([0, 1], cancer_label_names, rotation=20, ha="right")
axes[1].set_yticks([0, 1], cancer_label_names)
for i in range(2):
    for j in range(2):
        axes[1].text(j, i, int(confusion_df.iloc[i, j]), ha="center", va="center", color="#0f172a", fontweight="bold")

preview_df = sample_result.copy()
preview_df["标签"] = [cancer_label_names[int(v)] for v in preview_df["真实"]]
preview_df["预测标签"] = [cancer_label_names[int(v)] for v in preview_df["预测"]]
bar_colors = np.where(preview_df["真实"].to_numpy() == preview_df["预测"].to_numpy(), "#16a34a", "#dc2626")
axes[2].barh(np.arange(len(preview_df)), preview_df["预测置信度"], color=bar_colors)
axes[2].set_yticks(np.arange(len(preview_df)), [f"{a}->{b}" for a, b in zip(preview_df["标签"], preview_df["预测标签"])])
axes[2].invert_yaxis()
axes[2].set_xlim(0, 1)
axes[2].set_title("样本预测置信度", loc="left", fontweight="bold")
axes[2].grid(True, axis="x", color="#e2e8f0", linewidth=0.8)

plt.tight_layout()
plt.show()