The Shapley value is a concept from cooperative game theory that measures the amount of contribution from each player to the collective utility. Recent research attempts to use Shapley values in data valuation in the ML setting: evaluating the contribution of each training data point in improving the model performance. In this project, we test how robust the Shapley values are with respect to different training settings. We also devise methods to use the Shapley values to detect if the dataset contains mislabeled data or if the model is overfitting/underfitting on a specific dataset.
See full report for further details.