Examples

Data valuation

• Shapley values

  ---

  An introduction using the spotify dataset, showcasing grouped datasets and applied to improving model performance and identifying bogus data.

  

• KNN Shapley

  ---

  A showcase of a fast model-specific valuation method using the iris dataset.

  

• Data utility learning

  ---

  Learning a utility function from a few evaluations and using it to estimate the value of the remaining data.

  

• Least Core

  ---

  An alternative solution concept from game theory, illustrated on a classification problem.

  

• Data OOB

  ---

  A different and fast strategy for data valuation, using the out-of-bag error of a bagging model.

  

• Faster Banzhaf values

  ---

  Using Banzhaf values to estimate the value of data points in MNIST, and evaluating convergence speed of MSR.

  

Influence functions

• For CNNs

  ---

  Detecting corrupted labels with influence functions on the ImageNet dataset.

  

• For language models

  ---

  Using the IMDB dataset for sentiment analysis and a fine-tuned BERT model.

  

• For mislabeled data

  ---

  Detecting corrupted labels using a synthetic dataset.

  

• For outlier detection

  ---

  Using the wine dataset