KShapes, Resumed because of empty cluster

[pnwork]

Describe the bug
KShapes outputs "Resumed because of empty cluster", no clusters found; kmeans finds clusters

To Reproduce
default code with my dataset

Expected behavior
few clusters found

Environment (please complete the following information):

• OS: windows 10
• tslearn version 0.5.2

[kushalkolar]

it's been a while since I've used kshape clustering but this is probably dataset specific, you could try iterating through different seeds for the centroids

[AkievaAlina]

I have the same problem
kshape returns empty clusters for almost every random state
Increasing the number of n_init does not help
The problem is in the algorithm

[charavelg]

What kind of initialization (kshape init parameter) do you use?

For random initialization, we should make sure we don't select the same observation multiple times.
You can try replacing

tslearn/tslearn/clustering/kshape.py

Line 161 in 3165f09

indices = rs.choice(X.shape[0], self.n_clusters)

with :
indices = rs.choice(X.shape[0], self.n_clusters, replace=False)

Even though you may still encounter some "empty cluster", it will decrease the probability of such behavior.

[AkievaAlina]

What kind of initialization (kshape init parameter) do you use?

For random initialization, we should make sure we don't select the same observation multiple times. You can try replacing

tslearn/tslearn/clustering/kshape.py

Line 161 in 3165f09

indices = rs.choice(X.shape[0], self.n_clusters)
with :
indices = rs.choice(X.shape[0], self.n_clusters, replace=False)
Even though you may still encounter some "empty cluster", it will decrease the probability of such behavior.

Thank you for the respond!
I use the default parameter for initialization ('random')
What you suggested did not help. I still get empty clusters.
The numer of clusters is much smaller comparing to the number of time series (n_ts)

[charavelg]

#385 and #596 can probably help.

On top of this, you could check for "duplicates" in your dataset (which may lead to empty clusters) with https://github.com/tslearn-team/tslearn/blob/6bcf978ab615c97188e39eeecb05de17dc077ee0/tslearn/metrics/cycc.py#L54C5-L54C24
Something like
1 - cdist_normalized_cc(X, X, numpy.linalg.matrix_norm(X), numpy.linalg.matrix_norm(X), False)
should give you the SBD distance matrix where 0 means perfect similarity (or "duplicates").

Also note that data should be preprocessed with TimeSeriesScalerMeanVariance as shown in here.

Hope this helps!

[charavelg]

Is your dataset multivariate, in which case you might face #288 ?