Beyond Adult and COMPAS: Fairness in Multi-Class Prediction

Wael Alghamdi, Hsiang Hsu, Haewon Jeong, Hao Wang, P. Winston Michalak, Shahab Asoodeh, Flávio P. CalmonarXiv 2022 · DOI · OpenReview

TL;DR

This paper introduces FairProjection, a novel post-processing method for making multi-class machine learning classifiers fair across different groups. It tackles the limitations of existing fairness methods by being model-agnostic, scalable to large datasets, and capable of handling multiple protected groups and classes simultaneously.

What Problem Does This Paper Address?

Many real-world machine learning applications, such as loan applications, hiring, and healthcare, involve making decisions that have significant social consequences. Ensuring fairness in these decisions is crucial to prevent discrimination. While much research has focused on fairness in binary classification (e.g., approve/reject), many real-world scenarios involve multi-class predictions (e.g., grading students, disease severity). Existing fairness interventions are often tailored to binary outcomes or binary groups and are typically evaluated on small, overused datasets. This paper addresses the need for a versatile, scalable, and effective fairness intervention method for multi-class classification tasks that can handle multiple protected groups.

Methods

The core of FairProjection is an information-projection approach. The authors frame fairness as "projecting" a pre-trained, potentially unfair classifier onto the set of models that satisfy target group-fairness requirements. This projection is achieved by post-processing the classifier's outputs with a multiplicative factor.

Key methodological contributions include:

  • Information Projection Formulation: They adapt information projection, a concept from information theory, to the fairness context.
  • Finite-Sample Approximation: They develop an efficient algorithm, FairProjection, to compute the projected classifier using finite data samples, overcoming the limitation of prior work that assumed known probability distributions.
  • Parallelizable Iterative Algorithm: FairProjection employs an ADMM-based algorithm that is highly parallelizable, enabling it to scale to very large datasets.
  • Theoretical Guarantees: The authors provide theoretical guarantees for their algorithm in terms of sample complexity and convergence rates.
  • General f-divergence Support: The method is designed to work with various f-divergences, allowing flexibility in defining the "distance" between distributions.

Datasets

The authors evaluate FairProjection on several datasets:

  • UCI Adult: A commonly used dataset for fairness research.
  • COMPAS: Another widely used dataset for fairness research, particularly in criminal justice.
  • HSLS (High School Longitudinal Study): A dataset from the education domain.
  • ENEM (Exame Nacional do Ensino Médio): A newly introduced, large-scale dataset from Brazil's national high school exam. This dataset is significant because it contains over 1 million samples, multiple classes, and multiple intersectional protected groups, addressing the authors' call for larger and more diverse benchmarks beyond the traditional Adult and COMPAS datasets.

Key Insights and Ideas

  • Bridging the Gap to Multi-Class Fairness: The paper makes a significant contribution by extending fairness interventions beyond binary classification, a crucial step for real-world applicability.
  • Scalability is Key: The development of a parallelizable algorithm is essential for applying fairness methods to the massive datasets common in modern machine learning.
  • Post-Processing Power: FairProjection demonstrates that effective fairness interventions can be achieved through post-processing, without needing to retrain the original classifier from scratch.
  • ENEM as a New Benchmark: The introduction of the ENEM dataset provides a valuable resource for the fair ML community, encouraging research on more complex and realistic scenarios.

Key Results

  • Competitive Performance: FairProjection achieves competitive performance in terms of accuracy-fairness trade-offs, often outperforming or matching state-of-the-art methods.
  • Significant Runtime Improvements: The parallelizable nature of FairProjection leads to substantial speedups, especially on large datasets like ENEM. For instance, parallelization achieved a 15x speedup on the ENEM dataset.
  • Effective Multi-Class and Multi-Group Fairness: The method successfully reduces fairness violations (e.g., MEO, statistical parity) with minimal accuracy loss across various multi-class and multi-group settings.
  • Robustness: FairProjection shows consistent performance across different base classifiers (Logistic Regression, Random Forest, GBM) and datasets.

Limitations and Future Work

  • Group Fairness Focus: The current formulation primarily addresses group fairness. The authors suggest incorporating other fairness notions, such as individual fairness, as a future research direction.
  • Influence of Base Classifier Accuracy: The paper acknowledges that understanding how the accuracy of the pre-trained (base) classifier influences the performance of the projected classifier is an area for future investigation.
  • Data Availability Constraints: The performance of FairProjection, like any fairness method, is inherently constrained by the availability and quality of data. Performance might degrade with an increasing number of groups, labels, or fairness constraints.
  • Assumption 1: While the authors state that Assumption 1 (related to the properties of the f-divergence and the base classifier) is not a strict limitation in practice, it simplifies theoretical derivations. Further work could explore relaxing these assumptions.

Summary updated

Abstract

We consider the problem of producing fair probabilistic classifiers for multi-class classification tasks. We formulate this problem in terms of "projecting" a pre-trained (and potentially unfair) classifier onto the set of models that satisfy target group-fairness requirements. The new, projected model is given by post-processing the outputs of the pre-trained classifier by a multiplicative factor. We provide a parallelizable iterative algorithm for computing the projected classifier and derive both sample complexity and convergence guarantees. Comprehensive numerical comparisons with state-of-the-art benchmarks demonstrate that our approach maintains competitive performance in terms of accuracy-fairness trade-off curves, while achieving favorable runtime on large datasets. We also evaluate our method at scale on an open dataset with multiple classes, multiple intersectional protected groups, and over 1M samples.
BibTeX
@misc{alghamdi2022beyond,
  author = {Wael Alghamdi and Hsiang Hsu and Haewon Jeong and Hao Wang and P. Winston Michalak and Shahab Asoodeh and Flávio P. Calmon},
  title = {{Beyond Adult and COMPAS: Fairness in Multi-Class Prediction}},
  year = {2022},
  eprint = {2206.07801},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2206.07801},
}
Figure