Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Beyond Predictions: A Participatory Framework for Multi-Stakeholder Decision-Making

Vittoria Vineis, Giuseppe Perelli, Gabriele Tolomei

TL;DR

This work reframes automated decision-making from a sole focus on predictive accuracy to a multi-stakeholder paradigm that learns stakeholder preferences and selects actions via principled compromise rules. By integrating reward signals, surrogate preference models, outcome prediction, and a suite of aggregation operators within a single, modular pipeline, the framework enables context-aware, accountable recommendations across high-stakes domains. It unifies concepts from welfare economics, game theory, and multi-objective optimization, and introduces a general robustness certificate to quantify decisions' sensitivity to perturbations in stakeholder rewards. Empirical results in lending and healthcare illustrate improved trade-offs between predictive performance, fairness, and welfare, while robustness analysis ensures resilience to adversarial or noisy preferences. The approach offers a scalable, model-agnostic path toward participatory AI that emphasizes transparency, traceability, and real-world applicability in socially impactful deployments.

Abstract

Conventional automated decision-support systems often prioritize predictive accuracy, overlooking the complexities of real-world settings where stakeholders' preferences may diverge or conflict. This can lead to outcomes that disadvantage vulnerable groups and erode trust in algorithmic processes. Participatory AI approaches aim to address these issues but remain largely context-specific, limiting their broader applicability and scalability. To address these gaps, we propose a participatory framework that reframes decision-making as a multi-stakeholder learning and optimization problem. Our modular, model-agnostic approach builds on the standard machine learning training pipeline to fine-tune user-provided prediction models and evaluate decision strategies, including compromise functions that mediate stakeholder trade-offs. A synthetic scoring mechanism aggregates user-defined preferences across multiple metrics, ranking strategies and selecting an optimal decision-maker to generate actionable recommendations that jointly optimize performance, fairness, and domain-specific goals. Empirical validation on two high-stakes case studies demonstrates the versatility of the framework and its promise as a more accountable, context-aware alternative to prediction-centric pipelines for socially impactful deployments.

Beyond Predictions: A Participatory Framework for Multi-Stakeholder Decision-Making

TL;DR

This work reframes automated decision-making from a sole focus on predictive accuracy to a multi-stakeholder paradigm that learns stakeholder preferences and selects actions via principled compromise rules. By integrating reward signals, surrogate preference models, outcome prediction, and a suite of aggregation operators within a single, modular pipeline, the framework enables context-aware, accountable recommendations across high-stakes domains. It unifies concepts from welfare economics, game theory, and multi-objective optimization, and introduces a general robustness certificate to quantify decisions' sensitivity to perturbations in stakeholder rewards. Empirical results in lending and healthcare illustrate improved trade-offs between predictive performance, fairness, and welfare, while robustness analysis ensures resilience to adversarial or noisy preferences. The approach offers a scalable, model-agnostic path toward participatory AI that emphasizes transparency, traceability, and real-world applicability in socially impactful deployments.

Abstract

Conventional automated decision-support systems often prioritize predictive accuracy, overlooking the complexities of real-world settings where stakeholders' preferences may diverge or conflict. This can lead to outcomes that disadvantage vulnerable groups and erode trust in algorithmic processes. Participatory AI approaches aim to address these issues but remain largely context-specific, limiting their broader applicability and scalability. To address these gaps, we propose a participatory framework that reframes decision-making as a multi-stakeholder learning and optimization problem. Our modular, model-agnostic approach builds on the standard machine learning training pipeline to fine-tune user-provided prediction models and evaluate decision strategies, including compromise functions that mediate stakeholder trade-offs. A synthetic scoring mechanism aggregates user-defined preferences across multiple metrics, ranking strategies and selecting an optimal decision-maker to generate actionable recommendations that jointly optimize performance, fairness, and domain-specific goals. Empirical validation on two high-stakes case studies demonstrates the versatility of the framework and its promise as a more accountable, context-aware alternative to prediction-centric pipelines for socially impactful deployments.

Paper Structure

This paper contains 32 sections, 5 theorems, 30 equations, 1 figure, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Fairness and accountability in ADM
  4. Game-theoretic and social choice foundations.
  5. Multi-objective multi-agent systems
  6. Participatory AI as a paradigm.
  7. The Participatory Learning and Decision Framework
  8. Problem Formulation
  9. Core Framework Components
  10. Stakeholder preferences.
  11. Outcome Prediction Model.
  12. Decision strategies.
  13. Decision Evaluation.
  14. Best action selection.
  15. Computational Complexity
  16. ...and 17 more sections

Key Result

Theorem 1

Under Assumption assumption:discrete_action, relative to a baseline outcome-prediction system, the additional computational overhead introduced by the frameworks, added on top of the base cost, is: Offline (per cross-validation run): Online (per test instance): where $|G_q|$ is the size of the hyperparameter grid for stakeholder models $q_i$, $T_{\mathrm{val}}$ is the validation set size, $c_{\ma

Figures (1)

  • Figure 1: Comparison of decision strategy performance across evaluation metrics under different experimental conditions. Subfigures (a)--(c) report results for the lending scenario under variations in reward structure, predictive model, and sample size, respectively; subfigure (d) presents analogous results for the healthcare scenario. Each panel displays mean and standard deviation of test values for multiple compromise operators and baseline strategies, illustrating trade-offs between predictive accuracy, fairness, and welfare efficiency across configurations.

Theorems & Definitions (20)

  • Definition 1: Multi-Stakeholder Decision-Making Problem
  • Definition 2: Reward-augmented dataset
  • Definition 3: Actor-Specific Reward Function
  • Definition 4: Stakeholder Reward Prediction Model
  • Remark 1: Partial Reward Coverage and Generalization
  • Definition 5: Predicted Reward Matrix
  • Definition 6: Expected Reward Matrix
  • Definition 7: Decision Strategy
  • Definition 8: Compromise Function
  • Definition 9: Composite Evaluation Score
  • ...and 10 more