Please support custom distance function for k-means

[liufsd]

like : https://github.com/annoviko/pyclustering/blob/master/pyclustering/utils/metric.py

        # prepare input data and initial centers for cluster analysis using K-Means
        # create metric that will be used for clustering
        manhattan_metric = distance_metric(type_metric.MANHATTAN)
        # create instance of K-Means using specific distance metric:
        kmeans_instance = kmeans(sample, initial_centers, metric=manhattan_metric)
        # run cluster analysis and obtain results
        kmeans_instance.process()
        clusters = kmeans_instance.get_clusters()

from https://github.com/annoviko/pyclustering/blob/master/pyclustering/cluster/kmeans.py

[jnothman]

Mean of coordinates is only really appropriate for Euclidean distance. You want kmedoids or a specialised algorithm from this class. We have considered including kmedoids here but found it hard to identify a clear motivating example that set it as a better choice than agglomerative clustering, which supports custom metric (as do dbscan and optics).

[liufsd]

Maybe like this #12192 ?

[jnothman]

That case limits itself to cosine, which may be more applicable than other distances.

[liufsd]

Yes. i just need 'Mahalanobis Distance' , but i found it's no way~

[liufsd]

kmedoids :
https://github.com/scikit-learn/scikit-learn/pull/5085/files

[rth]

kmedoids

That work was moved to scikit-learn-contrib/scikit-learn-extra#12. It should be merged relatively soon.

[liufsd]

@rth wow~ thank you ~ ~ ~
but as i known 'waited a lot of time for this k-medoids algorithm'
so, i mean when it will be done ?

[rth]

so, i mean when it will be done ?

As soon as my few minor comments on that PR are addressed (if we don't wait for a second review that is).

[liufsd]

ok ok ~ please tell me when it's finish~ thanks~

[jnothman]

Please just subscribe to that issue and be quieter here. Thanks

[liufsd]

@jnhansen @rth
by the way, why not support Mahalanobis distance

PAIRED_DISTANCES = {
    'cosine': paired_cosine_distances,
    'euclidean': paired_euclidean_distances,
    'l2': paired_euclidean_distances,
    'l1': paired_manhattan_distances,
    'manhattan': paired_manhattan_distances,
    'cityblock': paired_manhattan_distances}

Mahalanobis distance for the data such as the user's weight and height. You known the euclidean_distances not support it.