AI-secure / Shapley-Study

[CVPR 2021] Scalability vs. Utility: Do We Have to Sacrifice One for the Other in Data Importance Quantification?
https://arxiv.org/abs/1911.07128v2
31 stars 3 forks source link

How to use as sorting algorithm #2

Open GeorgePearse opened 3 years ago

GeorgePearse commented 3 years ago

I'm now having some promising results with TMCShapley and KNNShapley. Thanks a lot for the good work. I'm looking for results that translate into incorrect label discovery and deep NN model performance improvements. So what I really care about is whether a datapoint has a positive value or a negative value.

Does anyone have any ideas for adaptations that fulfil this purpose, or other, none DataShapley techniques that already do something similar? e.g. greater accuracy for datapoints near a DataShapley value of 0 by disproportionately sampling from that set.

Thanks once more.