Fairness for the People, by the People: Minority Collective Action

TL;DR

The paper addresses fairness in ML by shifting focus from firm-side interventions to Algorithmic Collective Action (ACA), where a minority group relabels their own data to influence the learned model without altering the training pipeline. It introduces three model-agnostic methods (RB-label, RB-dist, RB-prob) to approximate counterfactual labels and demonstrates, across diverse datasets, that a small minority can substantially reduce fairness violations (e.g., EqOd) with limited accuracy loss. The work connects ACA to counterfactual and fair representation fairness, provides theoretical bounds on success and estimation error, and analyzes practical limitations when the minority is isolated. Overall, it shows a feasible, user-driven fairness mechanism with meaningful improvements and clear boundaries, suggesting directions for safer and more scalable deployment.

Abstract

Machine learning models often preserve biases present in training data, leading to unfair treatment of certain minority groups. Despite an array of existing firm-side bias mitigation techniques, they typically incur utility costs and require organizational buy-in. Recognizing that many models rely on user-contributed data, end-users can induce fairness through the framework of Algorithmic Collective Action, where a coordinated minority group strategically relabels its own data to enhance fairness, without altering the firm's training process. We propose three practical, model-agnostic methods to approximate ideal relabeling and validate them on real-world datasets. Our findings show that a subgroup of the minority can substantially reduce unfairness with a small impact on the overall prediction error.

Figures (15)

• Figure 1: Minority-only collective action can substantially improve fairness. With only $6$ label flips, the fairness violation of logistic regression goes down by over $75\%$ with only a negligible increase in prediction error. Circles and crosses represent majority and minority points, respectively.
• Figure 2: Visualization of KNN scoring methods with $k{=}3$. The minority is represented by the squares and the majority by circles, marked with a positive "$+$" or a negative "$-$" label. (a) RB-label: Two of the nearest majority neighbors have a positive label, resulting in the score $s{=}2$. (b) RB-dist: The average distance to the nearest positive majority neighbors results in the score $s{=}-\left(d_{1} {+} d_{2} {+} d_{3}\right)/3$.
• Figure 3: The lowest EqOd violation a collective can achieve greatly improves as the collective size increases, up to a certain point. Each point is a mean of 10 runs, with the standard deviation being smaller than the markers. In all the datasets we experimented on, the lowest EqOd violation converges around $\alpha=0.3$. Additional results are presented in \ref{['fig:sp_effect_alpha']} in the appendix.
• Figure 4: Our proposed methods are consistently more efficient than randomly flipping labels, requiring less label flips to attain the same level of EqOd. Each marker is the mean of 10 random runs with a specific number of label flips. The standard deviation is presented by the error bars. The dashed line shows the mean EqOd for a classifier trained on the dataset without collective action.
• Figure 5: Limiting the knowledge of the collective about the majority does not significantly harm the Pareto front. Each point is the mean of 10 runs and the curves are fitted to guide the eye.

Theorems & Definitions (20)

• Definition 1
• Proposition 1
• Proposition 2: Informal
• Proposition 3
• Proposition 4: Informal
• Definition 2: $\epsilon$-suboptimal classifier
• Proposition 4
• proof
• Proposition 5
• proof