Data Valuation

Data valuation in machine learning aims to assess the value of data or datasets for downstream tasks. This task is achieved by assigning a scalar to each element in the training set, reflecting its contribution to the final performance of a specific model, thereby enabling decision-making. Data valuation not only depends on specific models but also involves model-agnostic concepts such as cooperative game theory, neural network generalization estimates, and optimal transport theory. Its goal is to provide rigorous and practical methods for measuring data value, optimizing data markets, and the model training process.