Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Building Socially-Equitable Public Models

Yejia Liu, Jianyi Yang, Pengfei Li, Tongxin Li, Shaolei Ren

TL;DR

This paper tackles fairness for public models that support multiple downstream agents by introducing an Equitable Objective $\mathcal{L}_{EQ}^q$, inspired by $\alpha$-fairness, to promote uniform downstream performance. It develops Equitable PM, a policy-gradient–based training algorithm that handles both differentiable and non-differentiable downstream costs and can batch experiences to optimize across agents. The authors prove that increasing $q$ improves equity in the cost-regret distribution (via variance and entropy) and provide generalization bounds linking equitable training to out-of-sample performance. Empirically, two real-world–motivated cases (data-center carbon efficiency and EV charging) show that higher $q$ yields more uniform agent outcomes and lower average regret, even if pure prediction accuracy is not minimized. The work lays a foundation for socially responsible public models that balance diverse downstream objectives while highlighting privacy, robustness, and scalability as avenues for future work.

Abstract

Public models offer predictions to a variety of downstream tasks and have played a crucial role in various AI applications, showcasing their proficiency in accurate predictions. However, the exclusive emphasis on prediction accuracy may not align with the diverse end objectives of downstream agents. Recognizing the public model's predictions as a service, we advocate for integrating the objectives of downstream agents into the optimization process. Concretely, to address performance disparities and foster fairness among heterogeneous agents in training, we propose a novel Equitable Objective. This objective, coupled with a policy gradient algorithm, is crafted to train the public model to produce a more equitable/uniform performance distribution across downstream agents, each with their unique concerns. Both theoretical analysis and empirical case studies have proven the effectiveness of our method in advancing performance equity across diverse downstream agents utilizing the public model for their decision-making. Codes and datasets are released at https://github.com/Ren-Research/Socially-Equitable-Public-Models.

Building Socially-Equitable Public Models

TL;DR

This paper tackles fairness for public models that support multiple downstream agents by introducing an Equitable Objective , inspired by -fairness, to promote uniform downstream performance. It develops Equitable PM, a policy-gradient–based training algorithm that handles both differentiable and non-differentiable downstream costs and can batch experiences to optimize across agents. The authors prove that increasing improves equity in the cost-regret distribution (via variance and entropy) and provide generalization bounds linking equitable training to out-of-sample performance. Empirically, two real-world–motivated cases (data-center carbon efficiency and EV charging) show that higher yields more uniform agent outcomes and lower average regret, even if pure prediction accuracy is not minimized. The work lays a foundation for socially responsible public models that balance diverse downstream objectives while highlighting privacy, robustness, and scalability as avenues for future work.

Abstract

Public models offer predictions to a variety of downstream tasks and have played a crucial role in various AI applications, showcasing their proficiency in accurate predictions. However, the exclusive emphasis on prediction accuracy may not align with the diverse end objectives of downstream agents. Recognizing the public model's predictions as a service, we advocate for integrating the objectives of downstream agents into the optimization process. Concretely, to address performance disparities and foster fairness among heterogeneous agents in training, we propose a novel Equitable Objective. This objective, coupled with a policy gradient algorithm, is crafted to train the public model to produce a more equitable/uniform performance distribution across downstream agents, each with their unique concerns. Both theoretical analysis and empirical case studies have proven the effectiveness of our method in advancing performance equity across diverse downstream agents utilizing the public model for their decision-making. Codes and datasets are released at https://github.com/Ren-Research/Socially-Equitable-Public-Models.
Paper Structure (32 sections, 4 theorems, 43 equations, 6 figures, 4 tables, 1 algorithm)

This paper contains 32 sections, 4 theorems, 43 equations, 6 figures, 4 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Problem Formulation
  3. Fair Public Model for Downstream Agents
  4. Defining Fairness: An Equitable Objective
  5. Training Public Model: Equitable PM
  6. Theoretical Analysis
  7. Performance Equity/Uniformity
  8. Generalization Bounds
  9. Empirical Case Studies
  10. Evaluation Metrics
  11. Application I: Carbon Efficiency in Data Centers
  12. Datasets
  13. Implementation Details
  14. Results
  15. Application II: Scheduled EV Charging for Environmental Sustainability
  16. ...and 17 more sections

Key Result

Theorem 4.3

When $q=1$, the optimum of Equitable Objective is more equitable compared to $q=0$, indicated by smaller variance of the model performance distribution, i.e.$\mathsf{Var}(C_1(\theta^*_{q=1}), ..., C_M(\theta^*_{q=1})) < \mathsf{Var}(C_1(\theta^*_{q=0}), ..., C_M(\theta^*_{q=0}))$.

Figures (6)

  • Figure 1: The Equitable PM leads to a fairer solution by fostering a more equitable/uniform performance distribution across downstream agents. The embedded Equitable Objective $\mathcal{L}_{EQ}^q$ directly accounts for the decision costs across diverse downstream agents that use the prediction $\hat{y}$ from the public model $f$ for making informed decisions via agent-specific decision processes and actions $(\hat{a}_1, ..., \hat{a}_M)$.
  • Figure 2: (a) Comparison of cost regret distributions between Plain PM vs. Equitable PM on "similar agents" with different $\lambda$. The Equitable PM shows lower variability in cost distribution compared to the Plain PM. With varied $q+1$, we show cost regret distributions when using Equitable PM in (b) "similar agents" with different $\lambda$; (c) "different agents" with same $\lambda$; (d) "different agents" with different $\lambda$. As the value of q increases, the cost regret distribution across downstream agents achieves greater uniformity, implying a more equitable solution.
  • Figure 3: Statistics and cost regret distributions of test result between Plain PM and Equitable PM with varied $q+1$ for (a) "similar"; and (b) "different" agents. The Equitable PM demonstrates improved uniformity in agent distributions compared to the Plain PM. As $q$ increases, the uniformity of cost distribution across agents improves.
  • Figure 4: Depictions of (a) Azure workload demands shahrad2020serverless; (b) EV charging demands in ACN-Data acndata.
  • Figure 5: Depictions of home arrival, home departure and available charging time window for residential EV based on the NHTS government data NHTS.
  • ...and 1 more figures

Theorems & Definitions (10)

  • Definition 4.1
  • Definition 4.2
  • Theorem 4.3
  • Theorem 4.4
  • Proposition 4.5
  • proof
  • proof : Proof of the statement above
  • Corollary 1.1
  • proof
  • proof