Learning Time-Varying Convexifications of Multiple Fairness Measures
Quan Zhou, Jakub Marecek, Robert Shorten
TL;DR
This paper introduces a method for dynamically learning convexified fairness measures in machine learning models, accommodating multiple fairness notions with unknown and time-varying importance, using limited feedback.
Contribution
It proposes a novel approach to learn time-varying convexifications of multiple fairness measures with limited graph-structured feedback, addressing the challenge of unknown and changing fairness priorities.
Findings
Effective learning of fairness convexifications demonstrated
Handles multiple fairness notions simultaneously
Operates with limited feedback in complex settings
Abstract
There is an increasing appreciation that one may need to consider multiple measures of fairness, e.g., considering multiple group and individual fairness notions. The relative weights of the fairness regularisers are a priori unknown, may be time varying, and need to be learned on the fly. We consider the learning of time-varying convexifications of multiple fairness measures with limited graph-structured feedback.
Peer Reviews
Decision·Submitted to ICLR 2024
As demonstrated in the advertising example presented in the paper, considering the balance between various fairness regularizers in decision-making is crucial for devising a fair decision-making algorithm. Additionally, the setting where the weights between regularizers change over time is also particularly intriguing.
Although the paper is understandable, the structure/organization of the presentation could be improved. Specifically, since this paper introduces a new framework, I believe that the problem formulation (Section 3) and motivating example (Section 4) are crucial for understanding the paper. However, there are some unclear definitions and inconsistencies in the example. It is not clear as to what specific technical novelty exists beyond the proposal of a new framework. The description of how Alg
The paper addresses an interesting topic and Table 1 provides a comprehensive overview for earlier algorithms, which is helpful for readers who want to investigate further on this topic. Chapter 4 is an interesting read: motivating real-life examples are highly relevant with this topic (fairness) overall and may stimulate readers' interests to a great extent.
Graph-structured feedback is a fairly uncommon concept in the ML literature. Therefore, it would be immensely helpful if the paper has a more comprehensive section focusing on the preliminary materials. For example, what really are the "action vertices"? Does "action" refer to the concept in reinforcement learning, and even if so, what is its relationship with reducing unfairness here? (This is relevant with the previous comment above) Many equations are unclearly written and explained. Equatio
- The paper studies a significant and practical issue about integrating feedbacks from multiple fairness measures. In particular, these fairness measures could be conflicting to each other, or weighted differently at different times. I found the partial graph feedback framework proposed by the authors to be interesting and reasonable. - Overall the paper has a good clarity (except some typos, see below). The illustrative example was helpful in understanding how the OPT might look like and how th
- The exponentially-weighted algorithm only provides an upper bound for the weak regret, and there is few analysis / discussions on the optimality of such bound, and any upper / lower bounds analysis for the dynamic regret. - The numerical evaluations are on a relatively weak side. The fairness measures from the regularisers in the experiments are rather simple where each regulariser penalize the deviation from a fixed target share (essentially the same type of fairness measure). The results ca
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsEthics and Social Impacts of AI · Qualitative Comparative Analysis Research · Decision-Making and Behavioral Economics