Shapley-Value Data Valuation for Semi-supervised Learning

Semi-supervised learning aims at training accurate prediction models on labeled and unlabeled data. Its realization strongly depends on selecting pseudo-labeled data. The standard approach is to select instances based on the pseudo-label confidence values that they receive from the prediction models. In this paper we argue that this is an indirect approach w.r.t. the main goal of semi-supervised learning. Instead, we propose a direct approach that selects the pseudo-labeled instances based on their individual contributions for the performance of the prediction models. The individual instance contributions are computed as Shapley values w.r.t. characteristic functions related to the model performance. Experiments show that our approach outperforms the standard one when used in semi-supervised wrappers.