偏见

概述 AI 偏见是指机器学习系统在预测或决策中表现出系统性偏差，导致某些群体受到不公平对待。2025-2026 年，随着 AI 在招聘、信贷、医疗等高风险场景的广泛应用，偏见缓解已成为企业合规和伦理治理的核心议题。 一、偏见来源分析 1.1 数据层面偏见 偏见类型 描述 示例 历史偏见 训练数据反映历史歧视 招聘数据中女性占比过低 表示偏见 某些群体数据不足 医疗数据集中少数民族样本少 测量偏见 特征定义对某些群体不公平 用 zip code 代理信用评分 聚合偏见 模型对整体最优但对子群体不佳 整体准确率高但对老年群体差 1.2 模型层面偏见 偏见类型 描述 算法偏见 模型结构或损失函数强化偏见 部署偏见 模型在不同于训练环境的数据上运行 反馈偏见 模型输出影响后续数据收集，形成偏见循环 二、公平性定义与度量 2.1 数学定义 # 常见公平性指标计算 import numpy as np from sklearn.metrics import confusion_matrix def fairness_metrics(y_true, y_pred, sensitive_attr): """ 计算常见公平性指标 y_true: 真实标签 (0/1) y_pred: 预测标签 (0/1) sensitive_attr: 敏感属性 (0/1)，如性别、种族 """ group_0 = sensitive_attr == 0 group_1 = sensitive_attr == 1 # Demographic Parity Difference (DPD) # 要求不同群体的预测正例率相同 pred_positive_rate_0 = y_pred[group_0].mean() pred_positive_rate_1 = y_pred[group_1].mean() dpd = abs(pred_positive_rate_0 - pred_positive_rate_1) # Equal Opportunity Difference (EOD) # 要求不同群体的真正例率相同 tp_0 = y_pred[group_0 & (y_true == 1)].mean() tp_1 = y_pred[group_1 & (y_true == 1)].mean() eod = abs(tp_0 - tp_1) # Equalized Odds Difference (EQOD) # 要求不同群体的真正例率和假正例率都相同 fp_0 = y_pred[group_0 & (y_true == 0)].mean() fp_1 = y_pred[group_1 & (y_true == 0)].mean() eqod = max(abs(tp_0 - tp_1), abs(fp_0 - fp_1)) # Disparate Impact Ratio (DIR) dir_ratio = min(pred_positive_rate_0, pred_positive_rate_1) / \ max(pred_positive_rate_0, pred_positive_rate_1, 1e-10) return { "demographic_parity_diff": dpd, "equal_opportunity_diff": eod, "equalized_odds_diff": eqod, "disparate_impact_ratio": dir_ratio, "positive_rate_group_0": pred_positive_rate_0, "positive_rate_group_1": pred_positive_rate_1, } 2.2 指标对比 指标 定义 适用场景 法律依据 Demographic Parity P(Ŷ A=0) = P(Ŷ A=1) Equal Opportunity P(Ŷ=1 Y=1,A=0) = P(Ŷ=1 Y=1,A=1) Equalized Odds TPR 和 FPR 均相等 高风险决策 更严格 Disparate Impact P(Ŷ A=0) / P(Ŷ A=1) ∈ [0.8, 1.25] 三、数据层缓解技术 3.1 重采样（Resampling） from imblearn.over_sampling import SMOTE import pandas as pd def fairness_aware_oversample(df, sensitive_col, label_col, target_col): """ 基于公平性的过采样 确保每个敏感属性组合的样本比例均衡 """ # 计算每个组合的样本数 group_counts = df.groupby([sensitive_col, label_col]).size() target_count = group_counts.max() resampled_dfs = [] for (sensitive_val, label_val), group_df in df.groupby([sensitive_col, label_col]): if len(group_df) < target_count: # 对少数群体进行 SMOTE 过采样 smote = SMOTE(sampling_strategy=target_count, random_state=42) X = group_df.drop(columns=[label_col, sensitive_col]) y = group_df[label_col] X_resampled, y_resampled = smote.fit_resample(X, y) # 重新组装 DataFrame resampled_group = pd.concat([ X_resampled, pd.DataFrame({label_col: y_resampled}), pd.DataFrame({sensitive_col: [sensitive_val] * len(X_resampled)}) ], axis=1) resampled_dfs.append(resampled_group) else: resampled_dfs.append(group_df) return pd.concat(resampled_dfs, ignore_index=True) 3.2 重加权（Reweighting） def compute_fairness_weights(df, sensitive_col, label_col): """ 计算公平性感知的样本权重 思路：调整权重使得敏感属性与标签的联合分布均匀 """ n = len(df) group_sizes = df.groupby([sensitive_col, label_col]).size() weights = [] for _, row in df.iterrows(): s_val = row[sensitive_col] y_val = row[label_col] group_size = group_sizes[(s_val, y_val)] # 权重 = 总样本数 / (组数 * 组大小) n_groups = len(group_sizes) weight = n / (n_groups * group_size) weights.append(weight) return np.array(weights) # 在模型训练中使用 # sample_weight = compute_fairness_weights(train_df, "gender", "label") # model.fit(X_train, y_train, sample_weight=sample_weight) 四、训练层缓解技术 4.1 对抗性去偏（Adversarial Debiasing） import torch import torch.nn as nn class AdversarialDebiasing(nn.Module): """ 对抗性去偏模型 主预测器尝试预测目标 对抗器尝试从隐藏层预测敏感属性 """ def __init__(self, input_dim, hidden_dim, num_classes, num_sensitive): super().__init__() # 主预测器 self.encoder = nn.Sequential( nn.Linear(input_dim, hidden_dim), nn.ReLU(), nn.Linear(hidden_dim, hidden_dim), nn.ReLU(), ) self.predictor = nn.Linear(hidden_dim, num_classes) # 对抗器（用于学习去偏表示） self.adversary = nn.Sequential( nn.Linear(hidden_dim, hidden_dim // 2), nn.ReLU(), nn.Linear(hidden_dim // 2, num_sensitive), ) def forward(self, x, return_hidden=False): hidden = self.encoder(x) pred = self.predictor(hidden) if return_hidden: return pred, hidden return pred def adversary_pred(self, hidden): return self.adversary(hidden) def train_debiased_model(model, dataloader, sensitive_attr, epochs=10, lambda_adv=1.0): """ 训练对抗性去偏模型 lambda_adv: 对抗损失的权重（越大去偏越强） """ optimizer_pred = torch.optim.Adam( list(model.encoder.parameters()) + list(model.predictor.parameters()), lr=0.001 ) optimizer_adv = torch.optim.Adam(model.adversary.parameters(), lr=0.001) criterion_task = nn.CrossEntropyLoss() criterion_adv = nn.CrossEntropyLoss() for epoch in range(epochs): for batch_x, batch_y, batch_s in dataloader: # Step 1: 更新对抗器 optimizer_adv.zero_grad() _, hidden = model(batch_x, return_hidden=True) adv_pred = model.adversary_pred(hidden.detach()) loss_adv = criterion_adv(adv_pred, batch_s) loss_adv.backward() optimizer_adv.step() # Step 2: 更新主预测器（最小化任务损失，最大化对抗损失） optimizer_pred.zero_grad() pred, hidden = model(batch_x, return_hidden=True) loss_task = criterion_task(pred, batch_y) adv_pred = model.adversary_pred(hidden) # 梯度反转：最大化对抗器预测敏感属性的能力 # 即让隐藏表示对敏感属性"不可预测" loss_total = loss_task - lambda_adv * criterion_adv(adv_pred, batch_s) loss_total.backward() optimizer_pred.step() return model 4.2 公平性约束优化 # 使用 Fairlearn 进行公平性约束训练 from fairlearn.reductions import ExponentiatedGradient, DemographicParity from sklearn.linear_model import LogisticRegression def fair_train_with_constraint(X_train, y_train, sensitive_features, constraint="demographic_parity", eps=0.01): """ 使用 Fairlearn 的减少方法进行公平性约束训练 """ # 定义基础模型 base_model = LogisticRegression(solver='liblinear', fit_intercept=True) # 定义公平性约束 if constraint == "demographic_parity": fairness_constraint = DemographicParity() elif constraint == "equalized_odds": from fairlearn.reductions import EqualizedOdds fairness_constraint = EqualizedOdds() else: raise ValueError(f"不支持的约束类型: {constraint}") # 创建约束优化器 mitigator = ExponentiatedGradient( base_model, constraints=fairness_constraint, eps=eps # 容忍度 ) # 训练 mitigator.fit(X_train, y_train, sensitive_features=sensitive_features) return mitigator # 使用示例 # mitigator = fair_train_with_constraint( # X_train, y_train, sensitive_features, # constraint="demographic_parity", eps=0.01 # ) # y_pred_fair = mitigator.predict(X_test) 五、推理层缓解技术 5.1 后处理阈值调整 def calibrate_thresholds(y_proba, sensitive_attr, target_metric="demographic_parity"): """ 为不同敏感群体设置不同阈值以实现公平性 """ unique_groups = np.unique(sensitive_attr) thresholds = {} if target_metric == "demographic_parity": # 目标：使各群体的正例预测率相同 target_rate = y_proba.mean() # 全局正例率作为目标 for group in unique_groups: group_proba = y_proba[sensitive_attr == group] # 找到使该群体预测正例率达到目标的阈值 thresholds[group] = np.percentile(group_proba, (1 - target_rate) * 100) elif target_metric == "equalized_odds": # 需要迭代优化以满足 TPR 和 FPR 相等 pass # 实现略 return thresholds def fair_predict(y_proba, sensitive_attr, thresholds): """ 使用分组阈值进行预测 """ predictions = np.zeros_like(y_proba, dtype=int) for group, threshold in thresholds.items(): group_mask = sensitive_attr == group predictions[group_mask] = (y_proba[group_mask] >= threshold).astype(int) return predictions 5.2 模型集成去偏 def ensemble_fairness(models, X, sensitive_attr, strategy="voting"): """ 集成多个模型的预测结果以减少偏见 """ predictions = np.column_stack([m.predict(X) for m in models]) if strategy == "voting": # 简单投票 return predictions.mean(axis=1) >= 0.5 elif strategy == "fair_voting": # 公平性加权投票：为每个群体单独调整权重 final_pred = np.zeros(len(X)) for group in np.unique(sensitive_attr): group_mask = sensitive_attr == group group_pred = predictions[group_mask].mean(axis=1) final_pred[group_mask] = group_pred >= 0.5 return final_pred elif strategy == "stacking": # 使用元学习器学习如何组合（需额外训练） pass 六、评估与监控 6.1 公平性仪表盘 # 使用 AIF360 进行全面公平性评估 from aif360.metrics import BinaryLabelDatasetMetric, ClassificationMetric from aif360.datasets import BinaryLabelDataset def comprehensive_fairness_eval(dataset, predictions, sensitive_attr_idx=0): """ 全面的公平性评估 """ # 创建评估数据集 pred_dataset = dataset.copy() pred_dataset.labels = predictions # 二分类指标 metric = ClassificationMetric( dataset, pred_dataset, unprivileged_groups=[{sensitive_attr_idx: 0}], privileged_groups=[{sensitive_attr_idx: 1}] ) return { # 差异指标 "statistical_parity_diff": metric.statistical_parity_difference(), "disparate_impact": metric.disparate_impact(), "equal_opportunity_diff": metric.equal_opportunity_difference(), "average_odds_diff": metric.average_odds_difference(), # 绝对值指标 "false_positive_rate_diff": metric.difference( metric.false_positive_rate ), "false_negative_rate_diff": metric.difference( metric.false_negative_rate ), "accuracy_diff": metric.difference(metric.accuracy), # 整体性能 "overall_accuracy": metric.accuracy(), "balanced_accuracy": metric.accuracy( privileged=False ) * 0.5 + metric.accuracy(privileged=True) * 0.5, } 七、技术栈对比 层次 工具/库 核心功能 语言 数据 Fairlearn 公平性重采样、权重计算 Python 数据 AIF360 数据集变换、偏见检测 Python 训练 Fairlearn 约束优化、网格搜索 Python 训练 TensorFlow Fairness Indicators 模型评估 Python 推理 AIF360 后处理校准 Python 监控 Fairlearn Dashboard 可视化仪表盘 Python 监控 WhyLabs 生产环境监控 SaaS 参考 Barocas, S., et al. “Fairness and Machine Learning.” 2023. Bellamy, R., et al. “AI Fairness 360: An Extensible Toolkit for Detecting and Mitigating Algorithmic Bias.” IBM Journal of R&D, 2024. Fairlearn: https://fairlearn.org AIF360: https://aif360.res.ibm.com European Union AI Act (2024), Article 10-15 on Data Governance and Bias Mitigation. 加入讨论 这篇文章有姊妹讨论帖在硅基AGI论坛 — 全球首个碳基硅基认知交流平台。 ...