Inconsistent behaviour of selectors when n_select_features > rank(all_features) about scikit-matter HOT 3 CLOSED

Similar problem with FPS.

import numpy as np
from skcosmo.selection import FPS

X = np.ones((10,3))

n_features = 1
fs = FPS.FeatureFPS(X)
fs.select(n_features)
print("X.shape", X.shape)
print("n_features", n_features)
print("len(fs.idx)", len(fs.idx))
print("len(np.unique(fs.idx))", len(np.unique(fs.idx)))
print()

n_features = 3
fs = FPS.FeatureFPS(X)
fs.select(n_features)
print("X.shape", X.shape)
print("n_features", n_features)
print("len(fs.idx)", len(fs.idx))
print("len(np.unique(fs.idx))", len(np.unique(fs.idx)))

Output

X.shape (10, 3)
n_features 1
len(fs.idx) 1
len(np.unique(fs.idx)) 1

X.shape (10, 3)
n_features 3
len(fs.idx) 2
len(np.unique(fs.idx)) 2

FPS does only select features until the distance is close to zero and then stops. This could be FPS default behaviour, but it is really not clear from the python doc and the naming of the variables, and not consistent with CUR at the moment. I would rather have "select me exactly this amount, does not matter if these are crappy features" as default behaviour since PCA also does it this way and gives me an option to select the number of features in a smart way.

I also had the issue that one index was two times in fs.idx, but this was in an older version of skcosmo, so could have already been fixed

from scikit-matter.

Also there is somewhere an idx check wrong for FPS such that one index is included twice. I haven been able to reproduce this for a small script, but here a snippet of a larger part of the code

        fs = skcosmo.selection.FPS.FeatureFPS(features, tol=1e-50)
        fs.select(self.n_features)
        self.selected_idx_ = fs.idx
        print("len(self.selected_idx_)",len(self.selected_idx_))
        print("len(np.unique(self.selected_idx_))",len(np.unique(self.selected_idx_)))

Output

len(self.selected_idx_) 391
len(np.unique(self.selected_idx_)) 390

EDIT1: just checked, it is the first selected feature which is included twice in the idxs, one time at the beginning, and one time at the end. I think this is because at this point the distance is equal to 0, but an idx is added one last time because breaking from the for loop over n_select

EDIT2: Code example for one idx is included twice.

import numpy as np
from skcosmo.selection import FPS

X = np.array([[0, 100, 100, 100]])

n_features = 3
fs = FPS.FeatureFPS(X, idxs=[0])
fs.select(n_features)
print("X.shape", X.shape)
print("n_features", n_features)
print("len(fs.idx)", len(fs.idx))
print("len(np.unique(fs.idx))", len(np.unique(fs.idx)))

With output

X.shape (1, 4)
n_features 3
len(fs.idx) 3
len(np.unique(fs.idx)) 2

from scikit-matter.

Addressed in #58

from scikit-matter.