SphericalKMeans does not converge about spherecluster HOT 4 CLOSED

Hi @marcosterland,

In your example, ang is of shape ang.reshape((-1, 1)).shape = (300, 1) so it is a single vector there is not much to cluster.

The scipy.stats.vonmises routine only samples from a univariate vMF distribution. If you are trying to sample from a higher dimension mixture of vMF distribution, I recommend starting from this example: https://github.com/clara-labs/spherecluster/blob/develop/examples/small_mix_3d.py#L19-L36 and working with something like https://github.com/clara-labs/spherecluster/blob/develop/spherecluster/util.py .

cheers,
Jason

from spherecluster.

Actually I guess it's 300 vectors of length 1. In this case, normalizing each length-1 vector will result in the value 1 for each vector, making the exercise uneventful (which is why you get

[[ 1.]
 [ 1.]
 [ 1.]]

as output).

from spherecluster.

Thanks for your quick reply @jasonlaska .
It does make sense to cluster 1-D data. And in fact, the KMeans from scikit-learn finds the correct centers [[1.] [3.] [5.]] on the generated data.
But the scikit-learn KMeans uses standard Euclidean distance instead of the circular, so it's not applicable on e.g. angles.

from spherecluster.

I refer you to https://en.wikipedia.org/wiki/Cosine_similarity, the cosine between any two scalars is going to be same (as they all lie on the same axis, the angle between all of them is 0); similarly the inner product between then (just multiplication) normalized by their abs values (just multiplication) will always result in 1 or -1. Since the cosine distance/similarity does not take into account scale (as the euclidean distance does), it doesn't make sense to cluster scalars in this way unless they are complex scalars of the form a + b * i (and I'm not sure this package will handle that case).

I believe there might be a way to use the cosine distance in scikit-learn's k-means, you might want to give that a try.

from spherecluster.