PaperlayerView exact revision on arXiv

Source-linked AI summary

Towards Fair Classifiers Without Sensitive Attributes: Exploring Biases in Related Features

Tianxiang Zhao, Enyan Dai, Kai Shu, Suhang Wang

arXiv:2104.14537v4cs.LGcs.CY

TL;DR

Sensitive attributes are often unavailable to fairness methods, although correlated non-sensitive features may still reflect them. FairRF uses such features for prediction and as proxies for fairness regularization, with learned weights; the paper reports theoretical support and effective results on real-world datasets.

  • Problem

    Many fairness methods require sensitive attributes during training, but privacy, legal, or collection barriers can make them unavailable.

    Source: §1 · p. 1

  • Method

    FairRF uses correlated non-sensitive features both as classifier inputs and as pseudo sensitive attributes for prediction regularization, while learning feature-specific importance weights.

    Source: §1 · p. 2

  • Results

    Experiments on real-world datasets demonstrate the method’s effectiveness for fair classifiers with high classification accuracy.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The paper argues that minimizing prediction correlation with highly correlated related features can support fairness with respect to an unseen sensitive attribute.

    Source: §8 · p. 8

    Accurately specifying predefined importance weights can be difficult in real-world applications, motivating FairRF’s learned weighting approach.

    Source: §5.3 · p. 5

Abstract

from arXiv · show

Despite the rapid development and great success of machine learning models, extensive studies have exposed their disadvantage of inheriting latent discrimination and societal bias from the training data. This phenomenon hinders their adoption on high-stake applications. Thus, many efforts have been taken for developing fair machine learning models. Most of them require that sensitive attributes are available during training to learn fair models. However, in many real-world applications, it is usually infeasible to obtain the sensitive attributes due to privacy or legal issues, which challenges existing fair-ensuring strategies. Though the sensitive attribute of each data sample is unknown, we observe that there are usually some non-sensitive features in the training data that are highly correlated with sensitive attributes, which can be used to alleviate the bias. Therefore, in this paper, we study a novel problem of exploring features that are highly correlated with sensitive attributes for learning fair and accurate classifiers. We theoretically show that by minimizing the correlation between these related features and model prediction, we can learn a fair classifier. Based on this motivation, we propose a novel framework which simultaneously uses these related features for accurate prediction and enforces fairness. In addition, the model can dynamically adjust the regularization weight of each related feature to balance its contribution on model classification and fairness. Experimental results on real-world datasets demonstrate the effectiveness of the proposed model for learning fair models with high classification accuracy.

1 INTRODUCTION

Fairness methods often require sensitive attributes that are difficult to obtain, while models can inherit discrimination from training data. FairRF uses related non-sensitive features for prediction and fairness regularization, with learned weights intended to balance fairness and accuracy.

1 INTRODUCTION

  • Sensitive attributes are often unavailable because of privacy, legal, or data-collection barriers, limiting the use of existing fairness approaches.Earlier approaches preprocess data, regularize predictions, or post-process outputs, but require sensitive attributes during training.Source: §1 · p. 1
  • Non-sensitive features correlated with sensitive attributes can preserve useful classification information, so discarding them may reduce predictive value.The paper calls these features Related Features and notes that domain knowledge can help identify them for a target protected attribute.Source: §1 · p. 2
  • FairRF uses related features both as classifier inputs and as pseudo sensitive attributes for regularizing predictions.The framework is intended to learn fair and accurate classifiers without sensitive attributes.Source: §1 · p. 2
  • The paper theoretically argues that regularizing with related features can support fairness, and proposes automatically weighting each feature to balance fairness and accuracy.The weighting mechanism is also intended to address inaccurate or noisy identified related features.Source: §1 · p. 2
  • Experiments on real-world datasets demonstrate the proposed method’s effectiveness for fair classifiers with high classification accuracy.The introduction lists this as an experimental contribution.Source: §1 · p. 2

2 RELATED WORK

Prior fairness work includes individual, group, and Max-Min approaches, often requiring sensitive attributes or relying on proxies whose fit to protected groups can depend on data. FairRF instead studies related features for fairness without observing sensitive attributes.

2 RELATED WORK

  • Fairness methods are categorized as individual, group, or Max-Min fairness, while this paper focuses on group fairness.Group fairness aims to treat groups with different protected attributes equally.Source: §2 · p. 2
  • Pre-processing, in-processing, and post-processing methods modify data, training objectives, or predictions, respectively, to address fairness.The cited examples include revising attributes or labels, applying fairness constraints, and changing predictive labels.Source: §2 · p. 2
  • Many fairness methods require sensitive attributes; methods without them either target Max-Min fairness or use pseudo-groups whose conformity to protected groups depends on data distribution.Clustering-based groups may not correspond to the protected groups of interest.Source: §2 · p. 2
  • FairRF studies related features that correlate with unseen sensitive attributes and theoretically argues that regularizing predictions with them can support fairness.The paper says identifying these features requires little prior domain knowledge and reports that the supplied feature set can be incomplete or noisy.Source: §2 · p. 2

3 PROBLEM DEFINITION

The problem is to train a classifier that remains accurate and fair with respect to an unobserved protected attribute, using a predefined set of non-sensitive features correlated with that attribute. Those features provide alternative supervision when sensitive attributes are unavailable.

3 PROBLEM DEFINITION

  • When sensitive attributes are unavailable, correlated non-sensitive features can provide alternative supervision for learning a fair model while maintaining classification performance.The problem definition frames these features as a subset of the available feature set.Source: §3 · p. 3
  • The formal task is to learn a classifier with high accuracy and fairness on an unobserved protected attribute, given data, labels, and a predefined related-feature subset.Each related feature is specified as highly correlated with the protected attribute, such as race or gender.Source: §3 · p. 3
  • The setup assumes related features are supplied through domain knowledge or experts, although the set may be incomplete or noisy.FairRF is designed to reweight individual related features in response to this uncertainty.Source: §3 · p. 3

4 PRELIMINARY THEORETICAL ANALYSIS

The theoretical analysis relates Pearson correlations through angle bounds, then applies those bounds to explain why reducing prediction correlation with related features can support fairness toward an unseen sensitive attribute. It also extends the argument to multiple related features.

4 PRELIMINARY THEORETICAL ANALYSIS

  • Pearson correlation is used to measure linear correlation between variables.The coefficient is defined using centered variables and their standard deviations.Source: §4 · p. 3
  • The correlation-propagation result is presented as the theoretical justification for using related features to regularize predictions when the sensitive attribute is absent.The argument relies on Pearson correlation’s propagation property.Source: §4 · pp. 3–4
  • If correlations of X with Y and Y with Z are cos α and cos β, respectively, the correlation of X with Z lies in [cos(α + β), cos(α − β)].The analysis connects Pearson correlation to cosine similarity of vectors constructed from z-scores.Source: §4 · p. 3
  • When a related feature is highly correlated with the sensitive attribute, constraining its correlation with predictions near zero bounds the sensitive-attribute–prediction correlation.The paper states that when the feature-to-sensitive-attribute correlation approaches one and δ approaches zero, the prediction correlation with the sensitive attribute also approaches zero.Source: §4 · p. 4
  • For multiple related features, the sensitive-attribute–prediction correlation falls within the intersection of the ranges implied by each feature.The stated range is usually not tight, and greater divergence among related features can restrict it further.Source: §4 · p. 4

5 METHODOLOGY

FairRF uses related features as a surrogate fairness objective while a base classifier predicts labels, and learns feature-specific weights to balance fairness and predictive contribution. Its method targets binary classification and uses correlation regularization because sensitive attributes are unavailable.

5 METHODOLOGY

  • FairRF combines a base classifier, a covariance regularizer on related features, and a module that adjusts each feature’s importance score.The regularizer constrains correlation between related features and predictions as a surrogate fairness objective.Source: §5 · p. 4

5.1 Base Classifier

  • The backbone can be a neural network, logistic regression, or SVM; the method focuses on binary classification and leaves multiclass extension for future work.The classifier is trained with binary cross-entropy.Source: §5.1 · p. 4

5.2 Exploring Related Features for Fairness

  • When sensitive attributes are known, minimizing prediction correlation with them is a fairness objective; FairRF instead regularizes correlations with highly related non-sensitive features.The proxy regularization is intended to reduce prediction correlation with the unknown sensitive attribute.Source: §5.2 · p. 4
  • Feature-specific weights tune the proxy penalty: stronger weights are preferred for features more correlated with sensitive attributes, while weaker weights can limit noise in label prediction.Domain knowledge can help set the weights.Source: §5.2 · p. 5

5.3 Learning Importance of Related Features

  • FairRF learns weights from predefined initial estimates, constrains them to [0, 1], and adds an L2 penalty to discourage the trivial solution of assigning weight 1 to only the feature with the smallest regularization loss.Predefined weights and related-feature sets may be inaccurate, and selecting accurate initial weights can be difficult in real-world applications.Source: §5.3 · p. 5

6 OPTIMIZATION ALGORITHM

The optimization alternates between updating classifier parameters and related-feature weights. It uses gradient descent for classifier parameters and a constrained optimization procedure for the weights.

6 OPTIMIZATION ALGORITHM

  • With weights fixed, the method updates classifier parameters by gradient descent after removing objective terms irrelevant to those parameters.Alternating updates ease optimization of the constrained objective.Source: §6 · p. 5
  • With classifier parameters fixed, the weight update is treated as a convex primal problem and solved using KKT conditions.The conditions include stationarity, complementary slackness, and nonnegative weights.Source: §6 · p. 5
  • The weight procedure ranks regularization values in descending order, checks intervals to find a valid solution for v, and then calculates the weights.The full training pipeline is summarized in the supplementary Algorithm 1.Source: §6 · p. 5

7 EXPERIMENT

FairRF is evaluated on three benchmark datasets using accuracy and two group-fairness measures. The reported comparisons find improved fairness without significant accuracy loss, and test performance when related-feature sets are inaccurate or incomplete.

7.1 Datasets

  • The experiments use ADULT, COMPAS, and LSAC, with dataset-specific candidate related features selected using prior analysis or domain knowledge.The splits are 5:2:3 for train, evaluation, and test.Source: §7.1 · p. 6

7.2.1 Baselines.

  • Baselines include an unregularized vanilla model, a sensitive-attribute-aware reference, and methods that use pseudo-groups, remove candidate related features, or reweight regions.The authors note that ARL targets a different fairness formulation from the group fairness studied here.Source: §7.2.1 · p. 6

7.2.2 Configurations.

  • For the experiments, KSMOTE uses its provided code, while other approaches use a three-layer MLP with hidden dimensions 64 and 32 and Adam.The initial learning rate is 0.001.Source: §7.2.2 · p. 6

7.2.3 Evaluation Metrics.

  • The reported measures are accuracy, equal-opportunity difference, and demographic-parity difference; smaller fairness differences indicate fairer models.Equal opportunity concerns outcomes among positive instances, while demographic parity concerns positive rates across sensitive groups.Source: §7.2.3 · pp. 6–7

7.3 Classification Performance Comparison

  • For the main comparison, the base classifier is an MLP; hyperparameters are tuned on validation data, and results average five runs.FairRF uses dataset-specific β and η settings.Source: §7.3 · p. 7
  • FairRF improves both fairness measures over the vanilla model across all three datasets without significant performance drop, while other baselines show trade-offs.The reported comparison covers equal opportunity and demographic parity; Tables 1–3 are the dataset-specific comparison tables.Source: Table 1 · p. 6, Table 2 · p. 6, Table 3 · p. 6

7.4 Impact of the Quality of F𝑆on FairRF

  • FairRF remains effective with inaccurate related-feature sets, and automatic feature weighting performs better than fixed weights on fairness and accuracy.The reported quality-of-feature results cover ADULT and COMPAS; the authors say similar observations hold on LSAC.Source: §7.4 · p. 7, Table 4 · p. 7, Table 5 · p. 7

(a) ACC

FairRF’s related-feature weights mediate fairness–accuracy trade-offs, with sensitivity analysis and case studies illustrating their behavior. Experiments also report improved fairness with a small accuracy cost for a linear-regression backbone.

(a) ACC

  • Table 6 presents learned λ examples for selected related attributes alongside their correlations with the class label and sensitive attributes.Source: Table 6 · p. 8

7.5 Parameter Sensitivity Analysis

  • Larger η tends to improve fairness but can sharply reduce accuracy beyond a threshold; smaller β generally requires larger η to achieve fairness.Small β can make λ sparse, while β encourages a more uniform λ distribution and a faster accuracy drop as η increases.Source: §7.5 · p. 8

7.6 Case Study on 𝝀

  • FairRF gives higher weights to features highly correlated with S but weakly correlated with Y, and lower weights to label-predictive features, balancing fairness and accuracy.For example, Duration has correlation 0.02 with the label and 0.30 with S and receives weight 0.52; Relation receives weight 0.Source: §7.6 · p. 8

7.7 Flexibility of FairRF to Various Backbones

  • For LR, adding FairRF reduces ΔEO by 58.5% while accuracy drops by 2%.Source: §7.7 · p. 8

8 CONCLUSION

FairRF targets fair and accurate classification when sample-level sensitive attributes are unknown. It uses related features to regularize predictions and learns weights to balance fairness and accuracy; experiments report fairer performance with high accuracy.

8 CONCLUSION

  • FairRF addresses learning fair and accurate classifiers without knowing each sample’s sensitive attribute.Source: §8 · p. 8
  • The framework uses related features as pseudo-sensitive attributes to regularize model predictions.Source: §8 · p. 8
  • FairRF learns feature-importance weights to trade off each related feature’s contribution to classification accuracy and fairness.The paper reports fairer performance than existing approaches while maintaining high classification accuracy when sensitive attributes are unknown.Source: §8 · p. 8

A TRAINING ALGORITHM

FairRF’s training pipeline first pre-trains a model, then fine-tunes it for fairness with respect to related features. Related-feature weights can remain fixed or be updated iteratively.

A TRAINING ALGORITHM

B IMPLEMENTATION ON DIFFERENT BASE MODEL

FairRF is incorporated into LR and SVM as well as MLP, using gradient-based implementations. Across the evaluated models, fairness improves substantially with only a small accuracy decrease.

B IMPLEMENTATION ON DIFFERENT BASE MODEL

Related papers

Loading 2104.14537v4…