TL;DR
This paper introduces a novel framework for fairness in machine learning by defining fairness as low deviation in subgroup risks. This approach generalizes existing methods, handles complex sensitive features, and results in a convex optimization objective. By connecting to financial risk measures, particularly Conditional Value at Risk (CVaR), the authors provide a practical and theoretically sound method for learning fair classifiers.
What Problem Does This Paper Address?
Ensuring that machine learning models are fair, meaning they don't discriminate based on sensitive attributes like race or gender, is a critical challenge. Existing fairness definitions often struggle with:
- Limited scope of sensitive features: Many methods assume binary sensitive features (e.g., male/female), failing to handle categorical or continuous attributes.
- Non-convex optimization: Some fairness approaches lead to optimization problems that are difficult to solve efficiently and reliably.
- Lack of a unified framework: Different fairness notions can be hard to reconcile or generalize.
This paper aims to address these limitations by proposing a more general and mathematically tractable definition of fairness.
Methods
The core of the paper's contribution lies in a new definition of fairness based on subgroup risk deviation. Instead of just minimizing the average loss across all data points, the authors propose minimizing the deviation of losses across different subgroups defined by a sensitive feature.
Key methodological aspects include:
- Subgroup Risks: The loss (or risk) is first calculated for each subgroup induced by the sensitive feature.
- Risk Aggregation: These subgroup risks are then aggregated using a function that considers both their average and their deviation.
- Axiomatic Definition of Fairness Risk Measures: The authors axiomatize properties of "fairness risk measures" (R), which aggregate subgroup risks. These axioms are inspired by established risk measures in mathematical finance.
- Connection to Financial Risk Measures: A crucial insight is the strong connection between their fairness risk measures and established financial risk measures like Conditional Value at Risk (CVaR). This allows leveraging existing mathematical tools and understanding.
- Convex Optimization: By using CVaR as a specific risk aggregator, the paper derives a convex fairness-aware objective function. This is a significant advantage over methods that result in non-convex problems, making them easier to optimize.
- CVaR-based Objective: The proposed objective function, when using CVaR, effectively balances minimizing the average subgroup risk with ensuring that the risks across subgroups are commensurate (i.e., not too far apart).
Datasets
The paper uses the following datasets for empirical evaluation:
- Synthetic dataset (synth): A two-dimensional dataset.
- UCI Adult dataset: Used with gender as a binary sensitive feature and also with
fnlwgt(a continuous feature) as a sensitive attribute.
The paper does not introduce new benchmark datasets.
Key Insights and Ideas
- Fairness as Low Subgroup Risk Deviation: The most significant insight is reframing fairness not just as equal average outcomes, but as ensuring that the variability of outcomes across subgroups is low. This is a more nuanced and generalizable definition.
- Bridging Fairness and Finance: The paper elegantly connects the problem of algorithmic fairness to the well-developed field of financial risk management. This connection provides a rich theoretical foundation and practical tools for developing new fairness measures.
- CVaR as a Powerful Tool for Fairness: Using Conditional Value at Risk (CVaR) as a fairness risk measure leads to a convex optimization objective that naturally encourages commensurate subgroup risks. This offers a practical way to achieve fairness.
- Generalization to Complex Sensitive Features: The framework is designed to handle generic sensitive features (categorical, multi-valued, or even continuous), overcoming a key limitation of many prior works.
Key Results
The paper demonstrates the practical viability of their CVaR-based fairness approach through experiments on synthetic and real-world datasets.
- Fairness-Accuracy Tradeoff: As the parameter
α(controlling the focus on extreme subgroup risks in CVaR) is varied, the CVaR-based optimization method shows a clear tradeoff: increasingαgenerally leads to decreased predictive accuracy but also reduced fairness violation. - Competitive Performance: The CVaR-based approach is shown to be competitive with existing fair machine learning methods, such as the fair-ERM approach from Donini et al. [2018].
- Handling Continuous Sensitive Features: The paper shows that their framework can be applied to continuous sensitive features, leading to commensurate model predictions.
Limitations and Future Work
- Choice of Deviation Measure: While the paper axiomatizes desirable properties for deviation measures, the specific choice of measure (like standard deviation or CVaR) can influence the resulting fairness objective and its properties. The paper focuses on CVaR but suggests exploring other risk measures.
- Tuning Parameter
α: The CVaR-based objective involves a tuning parameterαthat controls the balance between average risk and risk deviation. Determining the optimalαfor a given application requires careful consideration. - Extensions to Other Fairness Notions: The paper primarily focuses on subgroup risk deviation. Future work could explore how this framework can be extended to other fairness notions or combined with them.
- Implications for Pre- and Post-processing: The paper focuses on direct optimization (type (c) in Section 2.4). Future work could explore implications for pre-processing and post-processing methods.
- Broader Exploration of Risk Measures: The paper suggests exploring other risk measures beyond CVaR, such as spectral measures, optimized certainty equivalents, and entropic value at risk, to further enrich the fairness toolkit.
Summary updated