PaperlayerView exact revision on arXiv

Source-linked AI summary

iFair: Learning Individually Fair Data Representations for Algorithmic Decision Making

Preethi Lahoti, Krishna P. Gummadi, Gerhard Weikum

arXiv:1806.01059v2cs.LGcs.IRstat.ML

TL;DR

Algorithmic fairness research has focused mainly on group-level guarantees, leaving individual fairness comparatively less explored. iFair learns probabilistic low-rank representations that preserve similarity on task-relevant attributes while supporting downstream classification and ranking. Across five real-world datasets, the authors report substantially more consistent decisions, with classification outperforming prior state-of-the-art work.

  • Problem

    Existing fairness research has prevalently pursued group fairness, while individual fairness has received relatively little attention despite its relevance to algorithmic ratings, rankings, and selections.

  • Method

    iFair probabilistically transforms records into low-rank representations that preserve fairness-aware similarity and utility across downstream classifiers and ranking tasks.

  • Results

    Across five real-world datasets, iFair produces substantially more consistent decisions; for classification, it outperforms state-of-the-art prior work.

  • Takeaways & Limitations

    Individual fairness can be incorporated as a dataset-level representation property while retaining applicability to classifiers and learning-to-rank tasks.

Abstract

from arXiv · show

People are rated and ranked, towards algorithmic decision making in an increasing number of applications, typically based on machine learning. Research on how to incorporate fairness into such tasks has prevalently pursued the paradigm of group fairness: giving adequate success rates to specifically protected groups. In contrast, the alternative paradigm of individual fairness has received relatively little attention, and this paper advances this less explored direction. The paper introduces a method for probabilistically mapping user records into a low-rank representation that reconciles individual fairness and the utility of classifiers and rankings in downstream applications. Our notion of individual fairness requires that users who are similar in all task-relevant attributes such as job qualification, and disregarding all potentially discriminating attributes such as gender, should have similar outcomes. We demonstrate the versatility of our method by applying it to classification and learning-to-rank tasks on a variety of real-world datasets. Our experiments show substantial improvements over the best prior work for this setting.

I. INTRODUCTION

Existing fairness approaches largely target protected groups and binary classification, while individual fairness must also preserve utility across broader decision tasks. iFair addresses this gap by learning flexible representations focused on individual fairness.

  • Motivation: Group-fairness methods commonly target statistical parity in binary classification and require explicit sensitive attributes or protected groups.
  • Motivation: Removing sensitive attributes before standard clustering can lose utility and miss correlations needed for individual fairness.
  • Limitations of Prior Work: Prior approaches are limited by binary-classification focus, use-case-specific representations, and objectives combining utility with group and individual fairness.
  • Approach: iFair learns a generalized representation preserving fairness-aware similarity while minimizing or bounding data loss, while intentionally excluding group fairness from its objective.
  • Contributions: iFair supports arbitrary classifiers and regression tasks, including ranking, and can handle multiple sensitive attributes whose critical values are selected at run time.
  • Contributions: The paper contributes an application-agnostic low-rank representation framework, individual fairness for learning-to-rank, and experiments covering classification and regression tasks.

II. RELATED WORK

Prior work emphasizes group fairness through statistical parity and related criteria, including in rankings, while individual fairness requires similar individuals to receive similar treatment. iFair instead defines fairness over representations using non-protected attributes.

  • Group Fairness: Group-fairness research includes statistical parity, equal odds, equal opportunity, calibration, and disparate mistreatment.
  • Individual Fairness: Individual fairness requires similar individuals to be treated similarly, formalized through distance-preserving probabilistic mappings over outcomes.
  • Fairness Methods: Fairness methods either preprocess data or impose fairness constraints within algorithmic models, while ranking work has primarily focused on group fairness.
  • iFair Setting: iFair assumes domain experts identify protected and non-protected attributes, but allows the protected-attribute set to be empty and protected values to be chosen downstream.
  • iFair Setting: Its individual-fairness definition requires records close on non-protected attributes to remain close after transformation, regardless of protected attributes.

A. Problem Formulation: Probabilistic Clustering

iFair learns a probabilistic, low-rank representation that balances reconstruction utility with preservation of similarities between records on non-protected attributes.

  • Probabilistic representation: Each record is represented through probabilities over K prototype vectors, producing a compact representation when K < M.The rows of U contain record-specific cluster probabilities, while the prototype vectors form V.
  • Probabilistic representation: The transformation mapping φ converts each input record xi into its learned representation ˜xi.
  • Utility objective: Utility is encouraged by minimizing reconstruction loss, defined as the sum of squared errors between X and ˜X.
  • Fairness objective: Individual fairness is encouraged by preserving pair-wise distances between records on non-protected attributes.
  • Combined objective: The final objective combines utility and fairness losses using hyper-parameters λ and µ.The combined loss is L = λ · Lutil(X, ˜X) + µ · Lfair(X, ˜X).

B. Probabilistic Prototype Learning

Prototype learning uses a flexible weighted distance and soft assignments to prototypes, with protected attributes intended to receive no or near-zero influence on similarity.

  • Distance and assignments: The distance function is chosen from Minkowski p-metrics, with p = 2 corresponding to a Gaussian kernel.
  • Distance and assignments: Softmax distances between records and prototypes define each record’s probability vector ui over the K clusters.The method avoids quadratic comparisons between all record pairs by comparing records only with prototypes.
  • Representation learning: The learned representation is obtained by mapping each record xi to ˜xi using the prototype probabilities and transformation φ.
  • Representation learning: Learning computes prototype vectors vk and attribute weights α that minimize the overall loss, with substitutions from the representation and distance equations.
  • Distance and assignments: The weight vector α controls each attribute’s influence, and protected attributes are assigned no or near-zero weight in the intended fairness setting.Experiments observe increased fairness when protected-attribute weights are near zero.

C. Gradient Descent Optimization:

Optimization minimizes the combined utility–fairness objective over prototype vectors and distance weights using L-BFGS.

  • Optimization objective: The system minimizes L = λ · Lutil(X, ˜X) + µ · Lfair(X, ˜X).
  • Learned parameters: The objective learns prototype vectors vk and the attribute-weight vector α as model parameters.
  • Optimization procedure: L-BFGS, a quasi-Newton algorithm, is applied to minimize Equation 10 and learn the model parameters.

IV. PROPERTIES OF THE iFair MODEL

The synthetic-data study examines whether iFair representations preserve utility while improving individual fairness. Compared with LFR, iFair is largely insensitive to protected-group changes and performs better on accuracy, consistency, and equality of opportunity, while LFR achieves better statistical parity.

  • Visualization: Figure 2 compares original, iFair, and LFR representations using class markers, protected-group colors, and classifier decision boundaries.Rows correspond to three synthetic cases, while learned positions in the middle and right columns are two-dimensional latent representations.
  • Main findings: iFair significantly outperforms LFR on accuracy, consistency, and equality of opportunity, whereas LFR wins on statistical parity.These metrics are reported for classifiers trained on the learned representations.
  • Influence of Protected Group: iFair representations remain largely unchanged when protected-group membership changes while all other attributes stay fixed.The learned representation and resulting decision outcome are therefore hardly influenced by the protected attribute in these synthetic cases.
  • Tension in Objective Function: LFR optimizes classifier accuracy, individual fairness, and statistical parity together, creating a pronounced tension with utility.The representations are stretched across the three objectives, ultimately sacrificing utility.
  • Tension in Objective Function: iFair optimizes utility and individual fairness while disregarding group fairness, making its multi-objective optimization more tractable than LFR’s.The study reports better decision boundaries and individual fairness for iFair than for LFR.
  • Trade-off between Utility and Individual Fairness: iFair’s improvement in individual fairness comes with a small utility drop caused by information loss when learning representative prototypes.The mapping function and pairwise distance function affect prototype learning, and alternative kernels or distances remain future-work directions.

V. EXPERIMENTS

The experiments test whether iFair can reconcile individual fairness and utility across both classification and learning-to-rank applications. The evaluation compares iFair with task-relevant baselines on these two use cases.

  • Experimental scope: The evaluation tests iFair’s ability to reconcile individual fairness and utility in classifier and learning-to-rank use cases.This assesses the versatility of an application-agnostic representation.
  • Baselines: The experiments compare iFair with LFR for classification and FA*IR for ranking, alongside a variety of baselines.The cited setup frames the comparisons around the two downstream task types.

A. Datasets

The study evaluates iFair on five publicly available datasets covering recidivism, income, credit risk, housing-listing rankings, and job-search rankings. Protected attributes and outcomes or ranking variables are selected for each domain.

  • Dataset collection: Five publicly available real-world datasets are used to evaluate iFair in settings drawn from algorithmic-fairness research.The datasets support both classification and ranking experiments.
  • Classification datasets: COMPAS uses race as the protected attribute and binary recidivism as the outcome.It is a widely used fairness test case for algorithmic decision making.
  • Classification datasets: Census Income uses gender as the protected attribute and income > 50K as the binary outcome.The dataset contains survey results for 48,842 US adults.
  • Classification datasets: German Credit uses age as the sensitive attribute and credit worthiness as the outcome across 1,000 credit-risk records.The attribute choices follow prior fairness literature.
  • Ranking datasets: Airbnb contains 27,597 cleaned records, with host gender as the protected attribute and rating/price as the ranking variable.Experiments use 22 informative categorical and numerical attributes from listings in five major US cities.
  • Ranking datasets: Xing provides the top 40 profiles for 57 German job queries, using gender as protected and a weighted experience, education, and profile-view score for ranking.Candidate information includes job category, work experience, education experience, profile views, and gender.
  • Scope and choices: The framework permits multiple protected attributes, multivariate outcomes, and inputs of all data types, while practical choices depend on domain experts and policies.The study’s dataset-specific choices are aligned with prior literature.

B. Setup and Baselines

The experiments compare iFair with representations and ranking methods under a common train-validation-test protocol. iFair is evaluated with alternative protected-attribute initializations, while mixture coefficients control utility and fairness trade-offs.

  • Protocol: Datasets are one-hot encoded and normalized to unit variance, then split into training, validation, and test sets shared by all methods.A grid search on the validation set selects hyperparameters.
  • Downstream models: The evaluation applies logistic regression for classification and linear regression for learning-to-rank comparisons.Both models operate on the evaluated data representations.
  • Baselines: Baselines include Full Data, Masked Data, SVD, SVD-masked, LFR, and FA*IR.FA*IR is a ranking method that returns a group-fair permutation without changing candidate scores.
  • iFair variants: iFair is tested with random attribute-weight initialization and initialization that sets protected-attribute weights near zero.The second initialization reflects discounting protected attributes in individual-fairness distance preservation while retaining numerical slack.
  • Training: Model parameters are randomly initialized unless specified otherwise, and each method-setting reports the best of three runs.This compensates for variation caused by parameter initialization.
  • Objective trade-offs: Mixture coefficients control trade-offs among utility, individual fairness, and applicable group fairness objectives.Hyperparameters are chosen under goals such as maximizing utility alone or combining utility with individual fairness.

C. Evaluation Measures

The evaluation measures quantify utility, individual fairness, and group fairness across classification and ranking settings. Classification experiments tune models along the utility–individual-fairness trade-off.

  • Utility is measured with accuracy and AUC for classification, and Kendall’s Tau and MAP for learning-to-rank.
  • Individual fairness is measured by outcome consistency between each individual and their 10 nearest neighbors.The metric captures whether similar individuals receive similar outcomes.
  • Group fairness is evaluated using Equality of Opportunity and statistical parity, with Equality of Opportunity as the primary metric.The measures compare outcomes or true-positive rates between protected and non-protected groups.
  • Classification models are selected from Pareto-optimal points with respect to AUC and individual-fairness consistency.Hyper-parameters are tuned via grid search.
  • iFair-b achieves the best overall classification trade-off, improving consistency while reducing accuracy by 6% and 10%, respectively, on Compas.For Census, the changes are +3.3% consistency and -7% accuracy; for Credit, +9% and -1.3%.

E. Evaluation on Learning-to-Rank Task

The learning-to-rank evaluation tests iFair on Xing and Airbnb rankings using utility, individual fairness, and group-fairness measures. iFair can be combined with FA*IR to impose target protected-group shares while retaining individual fairness.

  • Evaluation on Learning-to-Rank Task: Ranking utility is measured by Kendall’s Tau and average precision, while individual fairness uses consistency and group fairness uses the protected fraction in the top 10.The ranking experiments use multiple queries with corresponding ground-truth rankings.
  • Evaluation on Learning-to-Rank Task: Xing rankings use 57 job-search queries, while Airbnb rankings use 43 queries based on city, neighborhood, and home type.
  • Evaluation on Learning-to-Rank Task: Varying Xing score weights over [0.0, 0.25, 0.5, 0.75, 1.0] has no significant effect on the reported measures.The score combines work experience, education experience, and profile views.
  • Evaluation on Learning-to-Rank Task: FA*IR is the ranking baseline because it is designed to incorporate group fairness into rankings, whereas LFR is omitted because it is not geared toward regression tasks.
  • Evaluation on Learning-to-Rank Task: FA*IR’s utility in some configurations critically depends on its parameter p, so neither method is a clear overall winner under their different objectives.
  • Information Obfuscation & Relation to Group Fairness: The iFair representation substantially reduces adversarial protected-group prediction accuracy across datasets, despite correlated attributes.Masked Data retains enough implicit protected-group information for the adversarial classifier to perform fairly well.
  • Information Obfuscation & Relation to Group Fairness: Reduced protected-attribute information indirectly supports group fairness, with empirical improvements exceeding 10 percentage points despite group fairness not being an explicit goal.The extent of this benefit depends on the underlying data’s group base rates.
  • Information Obfuscation & Relation to Group Fairness: Combining iFair with FA*IR can achieve the required protected-group share while preserving the individual-fairness property of the learned representation.The ranking utility, protected share in the top 10, and individual fairness are evaluated as FA*IR’s parameter p increases.

VI. CONCLUSIONS

iFair transforms data into individually fair representations without requiring protected attributes, while supporting classifiers and learning-to-rank tasks. Across five real-world datasets, it substantially improves decision consistency, outperforms prior classification work, and reconciles utility with individual fairness to a large degree.

  • iFair is an unsupervised framework that probabilistically transforms data into individually fair representations for downstream applications.The representation-learning stage is application-agnostic and does not require access to protected attributes.
  • The framework supports both general classification and regression-based learning-to-rank tasks, with optional postprocessing for group fairness.Group criteria such as statistical parity can be enforced after learning the representations.
  • Across five real-world datasets, iFair reconciles utility and individual fairness to a large degree, producing substantially more consistent algorithmic decisions.The comparison is against decisions made on the original data.
  • For classification, iFair outperforms the state-of-the-art prior work and is presented as the first method to compute individually fair results for learning-to-rank tasks.
Loading 1806.01059v2…