Fast-DataShapley: Neural Modeling for Training Data Valuation

TL;DR

This paper tackles the problem of valuing training data for single-test-sample predictions using Shapley values, which are theoretically sound but computationally prohibitive as data providers grow. It introduces Fast-DataShapley, a one-pass, learned explainer that predicts training-data Shapley values for any new test sample without retraining the target model, leveraging a CWLS-style formulation with a prediction-specific value function. To reduce training overhead, it proposes three methods AFDS, GFDS, and GFDSPlus, offering theoretical guarantees and trade-offs between accuracy and efficiency; AFDS uses early-epoch information, GFDS groups data to reduce coalition size, and GFDSPlus adds symmetry-based grouping. Empirical results on MNIST and CIFAR-10 show substantial improvements in value estimation and dramatic reductions in training time, indicating strong practical potential for fair data compensation and copyright attribution in AI systems, with room for extension to generative tasks and privacy-preserving considerations.

Abstract

The value and copyright of training data are crucial in the artificial intelligence industry. Service platforms should protect data providers' legitimate rights and fairly reward them for their contributions. Shapley value, a potent tool for evaluating contributions, outperforms other methods in theory, but its computational overhead escalates exponentially with the number of data providers. Recent works based on Shapley values attempt to mitigate computation complexity by approximation algorithms. However, they need to retrain for each test sample, leading to intolerable costs. We propose Fast-DataShapley, a one-pass training method that leverages the weighted least squares characterization of the Shapley value to train a reusable explainer model with real-time reasoning speed. Given new test samples, no retraining is required to calculate the Shapley values of the training data. Additionally, we propose three methods with theoretical guarantees to reduce training overhead from two aspects: the approximate calculation of the utility function and the group calculation of the training data. We analyze time complexity to show the efficiency of our methods. The experimental evaluations on various image datasets demonstrate superior performance and efficiency compared to baselines. Specifically, the performance is improved to more than 2 times, and the explainer's training speed can be increased by two orders of magnitude.

Figures (6)

• Figure 1: An example of motivating AIGC data providers.
• Figure 2: The overview of the framework of return assignment.
• Figure 3: (a) The training of the explainer model mainly includes two core parts: the calculation of the value function and the other is updating the model through efficiency constraint. (b) Utilizing information from the first few epochs of sub-service model training to estimate the true value of utility. (c) A grouping function groups the training data, and each data of the same group and the overall data of other groups are regarded as players of the grand coalition to calculate the SV. (d) After grouping, each group is treated as a single player in the grand coalition.
• Figure 4: The comparison curve about value loss $H_\eta$ of different methods.
• Figure 5: The visualization of $AFDS$ on MNIST shows the test samples in the first column, with the corresponding true labels on the left. The next ten columns display the top 10 predictions from the explainer model that contributed most to the test samples. $o$ represents the original label, $p$ represents the predicted label of the service model, and $s$ represents the SV.

Theorems & Definitions (1)

• lemma 1