<ul><li>The value and copyright of training data are essential in the artificial intelligence sector, with service platforms needing to protect data providers' rights and reward them fairly.</li><li>Fast-DataShapley is introduced as a one-pass training method that uses the weighted least squares characterization of the Shapley value to create a reusable explainer model for real-time processing without the need for retraining for each test sample.</li><li>This method aims to address the computational complexities associated with traditional Shapley value computation by proposing three approaches to reduce training overhead, focusing on approximating the utility function and grouping the training data for more efficient calculations.</li><li>Experimental evaluations on various image datasets demonstrate significant improvements in performance and efficiency compared to existing methods, showcasing a more than 2.5 times performance enhancement and a two orders of magnitude increase in training speed for the explainer model.</li></ul>

Fast-DataShapley: Neural Modeling for Training Data Valuation

Discover more