Equality of Opportunity in Supervised Learning (1610.02413v1)

Abstract: We propose a criterion for discrimination against a specified sensitive attribute in supervised learning, where the goal is to predict some target based on available features. Assuming data about the predictor, target, and membership in the protected group are available, we show how to optimally adjust any learned predictor so as to remove discrimination according to our definition. Our framework also improves incentives by shifting the cost of poor classification from disadvantaged groups to the decision maker, who can respond by improving the classification accuracy. In line with other studies, our notion is oblivious: it depends only on the joint statistics of the predictor, the target and the protected attribute, but not on interpretation of individualfeatures. We study the inherent limits of defining and identifying biases based on such oblivious measures, outlining what can and cannot be inferred from different oblivious tests. We illustrate our notion using a case study of FICO credit scores.

An Overview of "Equality of Opportunity in Supervised Learning"

The paper "Equality of Opportunity in Supervised Learning" by Moritz Hardt, Eric Price, and Nathan Srebro puts forth a framework to address discrimination in supervised learning. It introduces the concept of ensuring fairness through a criterion termed as "equalized odds" and an associated relaxed version named "equal opportunity."

Abstract

The main contribution of this paper is the proposal of a new criterion for evaluating discrimination concerning sensitive attributes in supervised learning. The authors demonstrate how to adjust any learned predictor to eliminate discrimination, thus aligning incentivized fairness with better classification accuracy. They acknowledge the inherent limitation of defining and identifying biases using joint statistics and not individual feature interpretations.

Introduction

The introduction highlights the increasing role of machine learning in areas subject to anti-discrimination laws, such as credit scoring, employment, and criminal justice. It addresses the existing challenges and pitfalls in algorithmically ensuring fairness, referencing the Obama Administration's reports on ethical AI. The naive approach of ignoring protected attributes is declared insufficient due to the existence of redundant encodings. Similarly, demographic parity does not guarantee fairness and may lead to significant utility loss when target variables correlate with protected attributes.

Contributions

1. Fairness Criterion: The paper introduces the criterion of "equalized odds", which requires that predictor accuracy is equally high across all demographics.
2. Simplified Framework: It provides a practical method to derive non-discriminating predictors from any learned predictor through a post-processing step that does not alter the learning algorithm itself.
3. Theoretical Justification: The authors offer proofs that the Bayes optimal non-discriminating classifier can be obtained from the Bayes optimal regressor through their post-processing approach.
4. Constructive Limitations: The paper acknowledges the theoretical limitations of oblivious measures in distinguishing different dependency structures that may have varying fairness implications.

Equalized Odds and Equal Opportunity

The equalized odds criterion stipulates that the prediction and protected attribute should be conditionally independent. Specifically, true positive and false positive rates must be equal across protected groups. In a binary classification context, equal opportunity is a relaxed version that focuses only on true positive rates, ensuring that qualified individuals from all groups have similar chances of favorable outcomes.

Deriving Non-Discriminatory Predictors

The authors illustrate two primary methods for deriving fair predictors from existing models:

1. From Binary Predictors: They present a geometric approach to identify and achieve the optimal equalized odds classifier by transforming an arbitrary binary predictor through linear programming.
2. From Score Functions: For real-valued predictive scores, the paper provides a method to derive equalized odds predictors using threshold adjustments and potential randomization, which balance false and true positive rates across protected groups.

Case Study: FICO Scores

A substantial case paper on FICO scores, which predict creditworthiness, elucidates the practical implications of the proposed approach. By examining different classifier fairness constraints (e.g., max profit, race-blind, demographic parity, equal opportunity, and equalized odds), the paper demonstrates that equal opportunity-fair classifiers not only achieve significantly higher utility than demographic parity-based models but also better align fairness incentives by encouraging more accurate classification data collection.

Discussion and Conclusion

The paper advises that the proposed fairness measures should be used to identify and scrutinize potential discrimination rather than serving as definitive proofs of fairness. It emphasizes the alignment of fairness with the supervised learning objective of accuracy, thereby incentivizing the collection of better features and data to improve classifier performance across all demographic groups. The authors contend that the post-processing step should only be employed when data collection and feature enhancements have reached their limits, thereby transferring the burden of uncertainty from protected classes to decision-makers.

Implications and Future Research

The research introduces a rigorously defined fairness criterion, making substantial strides in addressing bias and discrimination in machine learning models. Future developments in this field could include more nuanced fairness notions and practical implementations, particularly for fields where predictive accuracy varies significantly across demographics. Additionally, prior to model deployment in sensitive applications, a robust validation of fairness criteria through empirical studies and domain-specific investigations will continue to be crucial.

This paper's framework and results are invaluable for advancing fairness in machine learning algorithms, ensuring balanced and equitable treatment across protected attributes while maintaining high predictive utility.