How are signatures computed ?

Here are the detailed steps to get the best features to distinguish two subsets (Either A vs B or A vs Rest):

1. For every feature, we compute the wasserstein distance between the 2 distributions (one for each subset).

2. We sort the features by the descending wasserstein distance and keep the 100 best features.

3. For each feature, we compute the cut with the best Matthews correlation coefficient to distinguish our two subsets,

4. Show the 20 features with the best MCC score.