Valda is a Python package for data valuation in machine learning. If you are interested in
- analyzing the contribution of individual training examples to the final classification performance, or
- identifying some noisy examples in the training set,
you may be interested in the functions provided by this package.
The current version supports five different data valuation methods. It supports all the classifiers from Sklearn for valuation, and also user-defined classifier using PyTorch.
- Leave-one-out (LOO),
- Data Shapley with the TMC algorithm (TMC-Shapley) from Ghorbani and Zou (2019),
- Beta Shapley from Kwon and Zou (2022)
- Class-wise Shapley (CS-Shapley) from Schoch et al. (2022)
- Influence Function (IF) from Koh and Liang (2017)
- IF only works with the classifiers built with PyTorch, because it requires gradient computation.
- The current version only support the first-order gradient computation, and we will add the second-order computation soon.
Please checkout a simple tutorial on Google Colab, for how to use this package.