含ADASYN的imblearn管道模型公平性优化报错求助

问题重现 #

我使用包含ADASYN的imblearn管道训练模型：

loaded_model = Pipeline(steps=[('feature_scaler', StandardScaler()), 
                ('adasyn_resampling', ADASYN(random_state=123)), 
                ('low_variance_remover', VarianceThreshold(threshold=0.01)), 
                ('feature_selection', SelectFromModel(estimator=RandomForestClassifier(random_state=0), max_features=35)), 
                ('classifier', CalibratedClassifierCV(base_estimator=LogisticRegression(C=0.1, penalty='l1', random_state=0, solver='saga'), method='isotonic'))])

尝试用Fairlearn的ExponentiatedGradient结合EqualizedOdds约束缓解模型不公平性：

from fairlearn.reductions import EqualizedOdds, ExponentiatedGradient

exponentiated_gradient = ExponentiatedGradient(
    estimator=loaded_model,
    constraints=EqualizedOdds(difference_bound=0.01), 
    sample_weight_name="classifier__sample_weight")

exponentiated_gradient.fit(X_train.drop(columns=['PCa-ID']), y_train, sensitive_features=sf_train)

运行时触发维度不匹配错误：

ValueError: sample_weight.shape == (4834,), expected (6967,)!

错误原因 #

错误核心是管道内的ADASYN重采样改变了样本数量：ExponentiatedGradient生成的sample_weight基于原始训练集（4834个样本），但ADASYN重采样后生成了更多样本（6967个），导致传递给最终分类器的样本权重维度与重采样后的数据集不匹配，触发校验错误。

解决方案 #

方案1：将重采样移到Fairlearn流程外（推荐）

先对原始训练集完成ADASYN重采样，再用处理后的数据集训练带公平约束的模型，从根源避免权重维度冲突：

from imblearn.over_sampling import ADASYN
from sklearn.pipeline import Pipeline

# 执行ADASYN重采样，同步处理敏感特征
adasyn = ADASYN(random_state=123)
X_resampled, y_resampled = adasyn.fit_resample(X_train.drop(columns=['PCa-ID']), y_train)
sf_resampled = sf_train.loc[adasyn.sample_indices_]

# 定义不含重采样步骤的管道
model_pipeline = Pipeline(steps=[
    ('feature_scaler', StandardScaler()), 
    ('low_variance_remover', VarianceThreshold(threshold=0.01)), 
    ('feature_selection', SelectFromModel(estimator=RandomForestClassifier(random_state=0), max_features=35)), 
    ('classifier', CalibratedClassifierCV(base_estimator=LogisticRegression(C=0.1, penalty='l1', random_state=0, solver='saga'), method='isotonic'))
])

# 训练公平模型
exponentiated_gradient = ExponentiatedGradient(
    estimator=model_pipeline,
    constraints=EqualizedOdds(difference_bound=0.01), 
    sample_weight_name="classifier__sample_weight")

exponentiated_gradient.fit(X_resampled, y_resampled, sensitive_features=sf_resampled)

方案2：自定义支持权重映射的重采样器

若必须在管道内保留ADASYN，可自定义采样器，将原始样本权重映射到重采样后的样本（少数类的副本继承原始样本的权重）：

from imblearn.base import BaseSampler

class WeightedADASYN(BaseSampler):
    def __init__(self, random_state=123):
        self.adasyn = ADASYN(random_state=random_state)
        self.sample_indices_ = None
        
    def fit_resample(self, X, y, sample_weight=None):
        X_resampled, y_resampled = self.adasyn.fit_resample(X, y)
        self.sample_indices_ = self.adasyn.sample_indices_
        
        # 将原始样本权重映射到重采样后的样本集合
        if sample_weight is not None:
            resampled_weights = sample_weight[self.sample_indices_]
            return X_resampled, y_resampled, resampled_weights
        return X_resampled, y_resampled

# 重新定义管道，使用自定义采样器
loaded_model = Pipeline(steps=[
    ('feature_scaler', StandardScaler()), 
    ('adasyn_resampling', WeightedADASYN(random_state=123)), 
    ('low_variance_remover', VarianceThreshold(threshold=0.01)), 
    ('feature_selection', SelectFromModel(estimator=RandomForestClassifier(random_state=0), max_features=35)), 
    ('classifier', CalibratedClassifierCV(base_estimator=LogisticRegression(C=0.1, penalty='l1', random_state=0, solver='saga'), method='isotonic'))
])

# 训练时传递顶层sample_weight，由自定义采样器处理映射
exponentiated_gradient = ExponentiatedGradient(
    estimator=loaded_model,
    constraints=EqualizedOdds(difference_bound=0.01), 
    sample_weight_name="sample_weight")

exponentiated_gradient.fit(X_train.drop(columns=['PCa-ID']), y_train, sensitive_features=sf_train)

内容的提问来源于stack exchange，提问作者Ana Rodrigues