Using Shapley Value for feature selection | Machine Learning

Shapley Values Explained: Explainability Models in Machine Learning | Chapter 03

We like to say “data is data” and that scale fixes everything. This skit questions that assumption using recent work on Data Shapley. By measuring how each training example actually affects validation loss during training, researchers found that some data helps, some is redundant, and some actively hurts performance. Removing negatively contributing data led to faster convergence with equal or better results. Paper: arXiv:2406.11011