for the same set of data, the centroids vary for new run

[krishnakumar85]

For each new run of node-kmeans on the same set of data, the clusters and centroids vary. Is there any way we can fix the skewed results or probably start with a constant seed.

[listonb]

I'm also seeing this problem. Appears to generate new centroids on every run of identical data

[Philmod]

Is there a lot of local minima in your data set?

[listonb]

Yes. This is pixel RGB color data from an image

[Philmod]

Yes, that's linked to your problem.

Finding the global minimum of the k-means problem is NP-hard in general.

[listonb]

Any easy fix?

[Philmod]

This is one todo.

I think that can be solved with different solutions:

• replicates: trying many random starting points and merging
• adding some randomness

I'm happy if you create a Pull Request with a solution.

Thanks, Philmod

[listonb]

Appreciate the time. I'll try to look into it after next week if i have some time!

[Morikko]

One of the solution used in sklearn is to used the inertia:

1. Do the kmean many times with different initiation
2. For each result, compute the inertia
3. Keep the results with the lowest inertia

Note about inertia (from sklearn): Sum of squared distances of samples to their closest cluster center.

@Philmod I can do a PR