Unlearning Shapley: Data Valuation through Machine Unlearning

ICLR 2026 Conference Submission17003 Authors

Keywords: data valuation, machine unlearning, shapley value

TL;DR: We propose Unlearning Shapley that computes Shapley values for full and partial data valuation through machine unlearning, enabling privacy-aware data valuation without retraining.

Abstract: Data valuation is essential for understanding and improving machine learning models. However, existing approaches such as Shapley-value-based retraining or influence functions are either computationally prohibitive or require full access to the training data, which is often unrealistic in practice. This challenge is particularly pressing in real-world settings such as data markets, federated environments, or compliance with the right to be forgotten, where only partial access to data subsets is available. We introduce Unlearning Shapley, a framework that adapts machine unlearning for both full-data and partial-data valuation. Instead of repeatedly retraining on all subsets, our method leverages a single pre-trained model and applies approximate unlearning to remove the effect of the target data, thereby estimating its marginal contribution. This design uniquely enables valuation when the rest of the training data is inaccessible, offering a privacy-compliant and practically deployable solution. Through theoretical analysis, we show the connection between Unlearning Shapley and classical Shapley values, and we provide bias and error bounds for our estimator. Experiments on benchmark vision datasets and large-scale language models demonstrate that Unlearning Shapley achieves comparable or superior performance to state-of-the-art methods in identifying influential or noisy data, while reliably extending to the partial-data setting where existing approaches fail. Our study highlights the importance of partial data valuation and extends the applicability of machine unlearning beyond privacy to equitable and transparent data markets.

Primary Area: interpretability and explainable AI

Submission Number: 17003