cs.LGAug 21, 2025
SaveMinority Collective Action for User-Side Fairness
Organizations: Max Planck Institute for Intelligent Systems, Tübingen AI Center, Tübingen, Germany · Department of Computer Science, University of Copenhagen · ETH Zurich
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 & tables
Figure 1 : Minority-only collective action can substantially improve fairness. When only 6 minority members change their labels, the EqOd violation ( Equation 2 ) of logistic regression drops by over with a negligible rise in prediction error ( Equation 1 ). Circles represent the majority and crosses represent the minority.
Figure 2 : Visualization of KNN scoring methods in Section 3 with . Squares represent the minority and circles the majority, 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 . ( b ) RB-dist: The average distance to the nearest positive majority neighbors results in the score .
Figure 3 : The lowest EqOd violation a collective can achieve decreases as the collective size increases, up to a certain point. Each point is the mean of 10 runs, with the standard deviation being smaller than the markers. Across the datasets we experimented on, the lowest EqOd violation stabilizes around . Additional results are presented in Figure 10 in the appendix.
Figure 4 : Our proposed methods are generally more efficient than randomly flipping labels, requiring fewer label flips to attain the same EqOd violation level. Each marker is the mean of 10 random runs with a specific number of label flips. Error bars show the standard deviation. The dashed line shows the mean EqOd for a classifier trained on the dataset without collective action.
Figure 5 : Limiting the collective’s knowledge of 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.
Figure 6 : The distribution used in proposition 5.1 . The color signifies the label, and the density shows the group membership.
Figure 7 : The user-side method cannot achieve zero EqOd violation, while the firm-side pre-processing method FARE [ 26 ] and the post-processing method calibrated equalized odds [ 41 ] attain 0 EqOd with large error. However, RB-prob has lower EqOd violation than the base classifier, with a smaller error than the firm-side methods.
Figure 8 : The Pareto fronts for using a fair representation when computing the KNN for RB-dist dominate the Pareto fronts for KNN computed on untransformed features. The blue stars represent the KNN without transforming the data, and the yellow triangles represent the KNN when the data is transformed using FARE [ 26 ] . The lines are fitted to guide the eye.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 9 : EqOd and SP violations per number of label flips for the random baseline, our method RB-prob, and the existing methods KDP [ 8 ] and CND [ 7 ] . Our method is more efficient than prior work, requiring fewer flips to achieve comparable violation levels. Note that in this experiment CND could flip any label, while all other methods were restricted to the labels of 30% of the minority.
Figure 10 : The lowest EqOd and SP violations a collective can achieve decrease 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, both violation metrics stabilize around .
Figure 11 : Our proposed methods are consistently more efficient than randomly flipping labels, requiring fewer label flips to attain comparable EqOd or SP violation levels. Each marker is the mean of 10 random runs with a specific number of label flips. The dashed line shows the corresponding mean violation for a classifier trained on the dataset without collective action.
Figure 12 : 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.
Figure 13 : The EqOd panel shows that the firm-side pre-processing method FARE [ 26 ] and the post-processing method calibrated equalized odds [ 41 ] attain 0 EqOd with large error. Across both EqOd and SP, RB-prob with (Section 3 ) has much smaller error and lower violations than the base classifier, but does not reach zero violation.
Figure 14 : Using a fair representation generally shifts the Pareto fronts lower and left across RB-prob, RB-label, and RB-dist. The blue stars represent each method without transforming the data, and the yellow triangles represent each method after transforming the data using FARE [ 26 ] . The lines are fitted by a polynomial of degree 2 to guide the eye.