search

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
30 stars 3 forks source link

Multi-label KNN Shapley #1

Open GeorgePearse opened 2 years ago

GeorgePearse commented 2 years ago

Very impressed by this piece of work. Has any thought gone into how the KNN Shapley approach could be adapted to the multi-label case while retaining its efficiency?

For that matter, what lead to the choice to only run a binary classification experiment between t-shirts and shirts on Fashion MNist, I would have thought that the method extends to multi-class without alteration.

ghost commented 2 years ago

Thanks for your question. Our approach is naturally applicable to the multi-label case. The reason why we ran the binary classification is to follow the experimental setting of the previous work Ghorbani et al. https://arxiv.org/abs/1904.02868.

GeorgePearse commented 2 years ago

@kkkkahlua sorry I believe you're mistaken. It works out-of-the-box on the multi-class case, e.g. MNist but not the multi-label case e.g. object detection. Please correct me again if I'm mistaken though. If I'm correct do you have any suggestions on how it may be adapted, or would this rule out any KNN based approaches? I know some KNN methods have an additional bayesian component that make it applicable to multi-label problems http://scikit.ml/api/skmultilearn.adapt.mlknn.html. I suspect it wouldn't be simple to implement this while maintaining the efficiency of your methods.