Using different metrics in k-medoids (sklearn), it does not work for all

[cmlp1957]

Hi,

First of all I would like to state that it is great to have k-medoids available in sk-learn. Many thanks for the effort done.

I have managed to use k-medoids from sk-learn only using ‘cosine’ as metric. When I use another metric I get the following type of message:

c:\users\cmlp\appdata\local\programs\python\python37\lib\site-packages\sklearn_extra\cluster\_k_medoids.py:235: UserWarning: Cluster 2 is empty! self.labels_[self.medoid_indices_[2]] may not be labeled with its corresponding cluster (2).
  "its corresponding cluster ({k}).".format(k=k)

I tried the following metrics:

metrics = ['braycurtis', 'canberra', 'chebyshev', 'cityblock', 'correlation', 'cosine', 'dice', 'euclidean',
'hamming', 'jaccard', 'kulsinski','matching', 'minkowski', 'rogerstanimoto',
'russellrao', 'seuclidean', 'sokalmichener', 'sokalsneath', 'sqeuclidean', 'yule']

But for some reason I am not able to generate the clusters using these other metrics. Since it is able to work swiftly for the 'cosine' metric I have the impression that the problem has not to do with the data. Where could than the problem be?

All the features I am using are categorical but two that are float. Before using k-medoids I used one-hot encoder for categorical features and minmax rescaling for the float variables.

Many thanks for your time,

Casiano Manrique

[rth]

Could you provide a minimal example? I have tried some of those metrics (including cityblock, canberra, minkowski, sqeuclidean) on plot_kmedoids_digits.py and they seem to be working fine.