search

marcotcr / lime

Lime: Explaining the predictions of any machine learning classifier
BSD 2-Clause "Simplified" License
11.59k stars 1.81k forks source link

Question about samples vs superpixels #534

Closed palatos closed 3 years ago

palatos commented 4 years ago

If I divide an image into a small number of superpixels/segments (let's say 3 or 4), the usual number of samples used for generating the explanation (say 200~1000) will probably mean the same combinations of perturbed segments will occur more than once, during the generation of the perturbed distribution.

Say two of my perturbed images have superpixels 1 and 3 turned off and the rest turned on.

Are these repeats a problem? Do they impact the explanation weights for each segment?

Thanks for any help/clarification on this.

marcotcr commented 3 years ago

They do impact the explanation weights, but I don't see why repeats would present any sort of problem. Their effect is similar to having two very similar perturbed samples with the same label.