Optimal Transport for Fairness: Archival Data Repair using Small Research Data Sets

Abigail Langbridge, Anthony Quinn, Robert ShortenIcdew 2024 · DOI

TL;DR

This paper introduces a novel method for repairing "unfair" data, specifically addressing the challenge of using limited labeled "research" data to fix large amounts of unlabeled "archival" data. Their approach leverages optimal transport to create fair data representations, demonstrating significant improvements in fairness metrics while maintaining data utility.

What Problem Does This Paper Address?

As Artificial Intelligence (AI) systems become more prevalent, ensuring their fairness is paramount. Regulations like the EU's AI Act are pushing for transparency and the mitigation of bias. A significant source of bias in AI models is the data they are trained on, which can reflect historical societal prejudices. While methods exist to "repair" unfair data, they often require large amounts of labeled data, which is expensive and time-consuming to obtain. This paper tackles the critical problem of repairing fairness in large, dynamic datasets where only a small fraction of the data is labeled with sensitive attributes (like race, gender, etc.). The challenge is to design a repair strategy using this limited labeled data that can then be efficiently applied to vast quantities of unlabeled archival data, which might be collected over time.

Methods

The core of the paper's contribution lies in its "distributional repair" method using optimal transport (OT). Instead of directly repairing individual data points, their approach focuses on learning the underlying data distributions.

Here's a breakdown of their key methodological innovations:

  • Fairness Definition: They define fairness as u-conditional independence, meaning that a protected attribute (S) should not be predictable from features (X) given an unprotected attribute (U). This is a stricter and more nuanced definition than simple overall independence.
  • Optimal Transport for Repair: They frame data repair as an optimal transport problem. The goal is to "transport" the unfair data distribution to a "fair" target distribution.
  • Barycentric Repair Target: They use a "fair barycenter" as the target distribution. This is a distribution that is equidistant (in a Wasserstein distance sense) from the distributions of the protected groups, effectively creating a balanced, fair representation.
  • Distributional Repair: Crucially, they propose a distributional repair rather than a point-wise repair. This means they learn the overall distribution of fair data and then use this to repair archival data. This is key for scalability.
  • Feature Stratification: To handle high-dimensional data, they stratify the repair process by individual features. This means they design separate OT plans for each feature, significantly reducing computational complexity.
  • Interpolated Supports: For archival data, they interpolate the marginal distributions from the small research dataset to create a support set. This allows the OT plans designed on the research data to be applied to the archival data.
  • Off-Sample Repair: Their method is designed to work on "off-sample" data, meaning data that was not used to design the repair. This is achieved by generalizing the OT plans learned on the research data to the archival data.

Datasets

  • Simulated Data: The paper uses simulated bivariate Gaussian sub-groups to validate their approach and study the impact of various operating conditions (research data size, interpolation resolution).
  • Adult Income Data Set [11]: This is a well-known benchmark dataset used to evaluate the performance of their method on real-world, noisy data. They restrict the feature space to "age" and "hours worked per week" for computational tractability.

Key Insights and Ideas

  • Small Labeled Data is Powerful: The research demonstrates that a surprisingly small amount of labeled "research" data can be sufficient to design effective fairness repair strategies for much larger "archival" datasets.
  • Distributional Repair is Scalable: Focusing on repairing distributions rather than individual points is crucial for handling the massive scale of archival data.
  • Feature-wise Repair is Efficient: Stratifying the repair by feature dramatically reduces the computational burden, making high-dimensional data repair feasible.
  • Generalization is Key: The ability of the repair strategy, learned on research data, to generalize to unseen archival data is a critical aspect of their method's success.
  • Fairness as a Distributional Property: The paper reinforces the idea that fairness is fundamentally a property of data distributions, and optimal transport provides a powerful mathematical framework to manipulate these distributions.

Key Results

  • Significant Reduction in Dependence: Their distributional OT-repair method significantly reduces the dependence of features on sensitive attributes, as measured by the $E_k$ metric (a form of symmetrized Kullback-Leibler divergence).
  • Comparable to State-of-the-Art (On-Sample): For on-sample repair (repairing data that was used to design the repair), their method performs comparably to existing geometric OT-based methods.
  • Superior for Off-Sample Repair: The key advantage is their method's ability to effectively repair off-sample (archival) data, a capability not offered by direct application of current methods.
  • Robustness to Data Size: The performance of their method converges even when the research data size is as low as 10% of the total data.
  • Efficiency Gains: The feature stratification and use of interpolated supports lead to significant computational savings, especially for large datasets. For instance, they show that a resolution ($n_Q$) of around 30 is sufficient for convergence.
  • Adult Income Data Performance: On the Adult Income dataset, they achieved substantial reductions in gender-education dependence, with a ~4-fold reduction in $E_k$ for research data and ~3-fold for archival data.

Limitations and Future Work

  • Assumption of Stationarity: The method relies on the assumption that archival data is drawn from the same stationary population as the research data. Real-world data often exhibits drift, which can impact repair performance.
  • Estimation of Sensitive Attributes: The method assumes that sensitive attributes for archival data are known or can be accurately estimated. In many practical scenarios, these attributes are unmeasured.
  • Feature Correlation: The feature-wise repair neglects intra-feature correlation structures, which could be a source of performance degradation.
  • Stopping Rules for Learning: Developing robust stopping rules for learning the marginal distributions is an open research question.
  • Non-Gaussian and Non-Continuous Features: Further work is needed to handle more complex data types and non-Gaussian, nonstationary features.
  • Trade-off between Repair and Information Loss: Investigating the trade-off between the degree of repair and potential information loss (data damage) is an area for future exploration.
  • Regularized OT and Monge Maps: Exploring the impact of regularized OT and the convergence to Monge maps on fairness metrics and individual fairness is also suggested.

Summary updated

BibTeX
@inproceedings{langbridge2024optimal,
  author = {Abigail Langbridge and Anthony Quinn and Robert Shorten},
  title = {{Optimal Transport for Fairness: Archival Data Repair using Small Research Data Sets}},
  booktitle = {Icdew},
  year = {2024},
  doi = {10.1109/icdew61823.2024.00036},
}
Figure