Research map / Learning theory and algorithms / Prediction and decision losses

Faithful Group Shapley Value

NeurIPS 2025 · Learning theory and algorithms › Prediction and decision losses · approach: Shapley value attribution

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Faithful Group Shapley Value (FGSV), the sum of individual Data Shapley values of group members, which resists shell company attacks, plus a fast approximation algorithm with a provable utility-evaluation complexity bound.

Problem

Group-level data valuation via Group Shapley Value (GSV) can be manipulated by splitting groups into subsidiaries (shell company attack), inflating a group's value without any change in data content.

Approach

Introduce a faithfulness axiom requiring a group's value to be invariant to how other data is partitioned; show that the unique method satisfying the axioms is the sum of individual Shapley values (FGSV), and approximate it directly by exploiting that only a few coefficient terms dominate, using paired Monte Carlo estimation of differences under a prudence/stability condition.

Key results

• On the Sum-of-Unanimity game with a 20,000 utility-evaluation budget, FGSV achieves the lowest AUCC and ARE and the lowest runtime per iteration among compared methods for n in {64,128,256}.
• Under a shell company attack in copyright attribution, GSV-based SRS inflates Google and Sprite shares, whereas FSRS remains consistent, avoiding misleading rankings (e.g., Google and Sprite outranking Vodafone under the Vodafone prompt).
• In the diabetes XAI experiment, GSV ranks age groups inconsistently across grouping schemes, while FGSV consistently ranks the Young group highest.

Measured effects reported in the paper

Surprising findings

• Under the prudence condition, splitting a group can increase its GSV, so standard GSV is manipulable by strategic splitting.

Limitations

• Theoretical guarantees rely on boundedness, second-order algorithmic stability, and O(1/s) deletion stability assumptions that may not hold for all utility functions.
• Does not address the copier attack; defense for general utility functions remains open.
• Small-subset utility values are undefined, requiring injection of non-informative data as an ad hoc remedy.

Experimental setup

Datasets: FlickrLogo-27 Diabetes dataset Sum-of-Unanimity cooperative game (synthetic)

Benchmarks: Sum-of-Unanimity cooperative game FlickrLogo-27 Diabetes dataset

Metrics: absolute relative error area under the convergence curve runtime per iteration

Baselines: Permutation estimator Group Testing estimator Complementary Contribution estimator One-for-All estimator KernelSHAP Unbiased KernelSHAP LeverageSHAP Group Shapley Value (GSV)

Builds on: Data Shapley Group Shapley Value Shapley Royalty Share LoRA KernelSHAP

Closest papers by meaning

• DU-Shapley: A Shapley Value Proxy for Efficient Dataset Valuation (NeurIPS 2024 · similarity 0.88)
• Rethinking Data Shapley for Data Selection Tasks: Misleads and Merits (ICML 2024 · similarity 0.84)
• Provably Accurate Shapley Value Estimation via Leverage Score Sampling (ICLR 2025 · similarity 0.84)
• shapiq: Shapley Interactions for Machine Learning (NeurIPS 2024 · similarity 0.84)
• Fair and Efficient Contribution Valuation for Vertical Federated Learning (ICLR 2024 · similarity 0.83)
• Generalized Priority-Aware Shapley Value (NeurIPS 2026 · similarity 0.83)