人工智能之核心基础 机器学习 第十二章 半监督学习

import numpy as np
from sklearn.datasets import fetch_20newsgroups
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.semi_supervised import LabelPropagation, LabelSpreading
from sklearn.linear_model import LogisticRegression
from sklearn.metrics import classification_report
from sklearn.model_selection import train_test_split

# 加载新闻数据
categories = ['alt.atheism', 'soc.religion.christian', 'comp.graphics', 'sci.med']
news = fetch_20newsgroups(subset='all', categories=categories, remove=('headers', 'footers'))

X_text = news.data
y_true = news.target

# 向量化
vectorizer = TfidfVectorizer(max_features=5000, stop_words='english')
X = vectorizer.fit_transform(X_text)

# 模拟：仅用1%数据有标签
np.random.seed(42)
n_total = len(y_true)
n_labeled = int(0.01 * n_total)  # 1%有标签

# 随机选择有标签样本
labeled_indices = np.random.choice(n_total, size=n_labeled, replace=False)
y_semi = np.full(n_total, -1)  # -1 表示无标签
y_semi[labeled_indices] = y_true[labeled_indices]

# 方法1：伪标签法（用逻辑回归）
lr = LogisticRegression(max_iter=1000)
lr.fit(X[labeled_indices], y_true[labeled_indices])

# 预测无标签数据，取高置信度（>0.95）
proba = lr.predict_proba(X)
pseudo_labels = lr.predict(X)
high_conf = np.max(proba, axis=1) > 0.95

# 构建新训练集
X_train_pseudo = X[np.concatenate([labeled_indices, np.where(high_conf)[0]])]
y_train_pseudo = np.concatenate([y_true[labeled_indices], pseudo_labels[high_conf]])

# 重新训练
lr_pseudo = LogisticRegression(max_iter=1000)
lr_pseudo.fit(X_train_pseudo, y_train_pseudo)

# 方法2：Label Spreading
ls = LabelSpreading(kernel='knn', n_neighbors=10, max_iter=100)
ls.fit(X, y_semi)

# 评估（用完整测试集）
test_indices = np.setdiff1d(np.arange(n_total), labeled_indices)
y_test = y_true[test_indices]

print("=== 伪标签法 ===")
print(classification_report(y_test, lr_pseudo.predict(X[test_indices])))

print("\n=== Label Spreading ===")
print(classification_report(y_test, ls.predict(X[test_indices])))

from sklearn.datasets import make_classification
from sklearn.semi_supervised import LabelSpreading
from sklearn.cluster import KMeans
from sklearn.metrics import adjusted_rand_score as ARI

# 模拟客户数据（收入、消费、活跃度）
X, y_true = make_classification(n_samples=1000, n_features=3, n_redundant=0, 
                               n_clusters_per_class=1, n_classes=3, random_state=42)

# 仅10个客户有标签（业务专家标注）
labeled_idx = np.random.choice(1000, size=10, replace=False)
y_semi = np.full(1000, -1)
y_semi[labeled_idx] = y_true[labeled_idx]

# 1. 纯无监督：K-Means
kmeans = KMeans(n_clusters=3, random_state=42)
y_km = kmeans.fit_predict(X)

# 2. 半监督：Label Spreading
ls = LabelSpreading(kernel='rbf', gamma=10)
ls.fit(X, y_semi)
y_ls = ls.predict(X)

# 评估（假设我们知道真实标签）
print("K-Means ARI:", ARI(y_true, y_km))
print("Label Spreading ARI:", ARI(y_true, y_ls))

# 可视化（前两维）
plt.figure(figsize=(12, 4))
plt.subplot(131)
plt.scatter(X[:, 0], X[:, 1], c=y_true, cmap='tab10')
plt.title("真实标签")

plt.subplot(132)
plt.scatter(X[:, 0], X[:, 1], c=y_km, cmap='tab10')
plt.title("K-Means (无监督)")

plt.subplot(133)
plt.scatter(X[:, 0], X[:, 1], c=y_ls, cmap='tab10')
plt.title("Label Spreading (半监督)")

plt.show()