[Paper Review] Collaborative Machine Learning with Incentive-Aware Model Rewards

Rachael Hwee Ling Sim, Yehong Zhang+2 more|arXiv (Cornell University)|Oct 24, 2020

Data Stream Mining Techniques33 citations

TL;DR

The paper proposes an incentive-aware model reward scheme for collaborative machine learning, valuing each party’s data with information gain and Shapley values, and realizing rewards by injecting Gaussian noise into aggregated data; it demonstrates properties and empirical performance on synthetic and real-world datasets.

ABSTRACT

Collaborative machine learning (ML) is an appealing paradigm to build high-quality ML models by training on the aggregated data from many parties. However, these parties are only willing to share their data when given enough incentives, such as a guaranteed fair reward based on their contributions. This motivates the need for measuring a party's contribution and designing an incentive-aware reward scheme accordingly. This paper proposes to value a party's reward based on Shapley value and information gain on model parameters given its data. Subsequently, we give each party a model as a reward. To formally incentivize the collaboration, we define some desirable properties (e.g., fairness and stability) which are inspired by cooperative game theory but adapted for our model reward that is uniquely freely replicable. Then, we propose a novel model reward scheme to satisfy fairness and trade off between the desirable properties via an adjustable parameter. The value of each party's model reward determined by our scheme is attained by injecting Gaussian noise to the aggregated training data with an optimized noise variance. We empirically demonstrate interesting properties of our scheme and evaluate its performance using synthetic and real-world datasets.

Motivation & Objective

• Motivate collaborative ML where participants require fair incentives to share data.
• Value each party’s data using information gain on model parameters to reflect informativeness.
• Design a model reward scheme that achieves fairness, stability, and group welfare under freely replicable rewards.
• Realize computed rewards by perturbing aggregated data with optimized Gaussian noise.
• Empirically evaluate the approach on synthetic and real-world datasets to illustrate properties and performance.

Proposed method

• Valuing data by information gain on model parameters given coalition data.
• Adapting cooperative game theory incentives to freely replicable model rewards (Shapley fairness, stability, individual rationality, group welfare).
• Introducing rho-Shapley fairness to trade off fairness and other incentives via an adjustable parameter rho.
• Mapping Shapley values to rewards with r_i = (phi_i/phi*)^rho * v_N to achieve rho-Shapley fairness.
• Realizing rewards by injecting Gaussian noise into other parties’ data and optimizing the noise variance eta_i to match target rewards.
• Empirical evaluation on synthetic and real datasets using Bayesian regression models to study information gain and reward behavior.

Experimental results

Research questions

• RQ1 How should data contributed by multiple parties be valued to incentivize collaboration?
• RQ2 Can a reward scheme ensure fairness, stability, and individual rationality in a collaboratively trained model scenario?
• RQ3 How can the decided model rewards be realized in practice without excessive data subset selection?
• RQ4 What is the effect of the rho parameter on fairness, group welfare, and stability?

Key findings

• The authors define a data valuation method via information gain on model parameters, avoiding dependence on a common validation dataset.
• A Shapley-based fairness framework is proposed and extended to rho-Shapley fairness to balance incentives; rewards scale with a constant k and exponent rho.
• The reward scheme satisfies non-negativity, feasibility, weak efficiency, fairness properties, and can promote grand-coalition stability under monotonic data value.
• Rewards are realized by adding Gaussian noise to other parties’ data and optimizing a continuous noise variance parameter to achieve target rewards.
• Empirical evaluations on synthetic Friedman data and real-world datasets (diabetes progression, California housing) demonstrate how information gain correlates with predictive performance and how rho affects IG and MNLP metrics.
• The approach highlights the potential to incentivize collaboration without monetary compensation by leveraging model rewards modeled as freely replicable digital goods.