Apparently, we don't have the equivalence dense-sparse anymore.

I deleted the redundant comments. They should no longer appear now that #23539 was merged: the description was edited by the bot 4 hours ago, so it seems to work as intended.

All failures appear with the DIAgonal sparse matrix format. The same assertions with CSR / CSC do seem to pass.

I will try to craft a minimal reproducer and report upstream.

The problem happens at fit time, the scale_ attribute of RobustScaler has different values for the DIAgonal input data:

dense [1.35  0.525 3.55  1.65 ]
csr [1.35  0.525 3.55  1.65 ]
csc [1.35  0.525 3.55  1.65 ]
bsr [1.35  0.525 3.55  1.65 ]
coo [1.35  0.525 3.55  1.65 ]
dia [1.4 0.7 3.6 1.6]

Ok, I found a minimal reproducer:

import numpy as np
import scipy.sparse as sp


X = np.ones(shape=(13, 4))
X[0, 0] = np.nan
print(sp.csc_matrix(X).indptr)
print(sp.csc_matrix(sp.dia_matrix(X)).indptr)

out:

[ 0 13 26 39 52]
[ 0 12 25 38 51]

I updated scipy/scipy#16365 with a more minimal reproducer.

FYI: The bug is actually in NumPy: numpy/numpy#21685

FYI: The bug is actually in NumPy: numpy/numpy#21685

Thanks. I should have checked I could reproduce with the latest stable numpy before reporting to scipy.

CI is no longer failing! ✅

Successful run

Closing this one, I am guessing this did not get closed automatically because auto-close is off by default since #23155

@thomasjpfan this bot is really cool :)