python import numpy as np from sklearn.ensemble import RandomForestClassifier from sklearn.neighbors import NearestNeighbors

def entropy(prob): return -np.sum(prob * np.log(prob + 1e-10), axis=1)

假设已有模型 model, 未标注池 X_pool, 每轮挑选数量 B

proba = model.predict_proba(X_pool) uncertainty = entropy(proba)

计算密度 (基于未标注池)

k = 10 nn = NearestNeighbors(n_neighbors=k, algorithm='auto').fit(X_pool) distances, _ = nn.kneighbors(X_pool)

平均距离的倒数作为密度，加上微小值防止除零

density = 1.0 / (np.mean(distances, axis=1) + 1e-5)

beta = 1.0 score = uncertainty * (density ** beta)

选择得分最高的B个样本索引

selected = np.argsort(score)[-B:]