MachineLearning.KMeans: Balanced clustering

[jagbarcelo]

Is there any tweak/configuration/settings/example that shows how to calculate clusters of same size using your Framework? If not, is this included in your plans for next releases?

Right now the only thing that I have managed to do is repeat the Learn method using different random seeds for the initial position of the centroid set, until (eventually) a balanced solution is found. Obviously this solution is not optimal at all. Despite the fact that the solutions found are not repeatable and the execution times vary.

Thanks.

[cesarsouza]

Hi there,

Right now the framework doesn't offer any mechanism to create clusters of the same size. However, this problem seems highly interesting. I will probably be adding a method for this task in the next few days.

Regards,
Cesar

[cesarsouza]

It seems in fact that a balanced k-means algorithm can be implemented using the Munkres / Hungarian algorithm for solving minimum-cost assignment problems, which would also be an excellent addition to the framework. I will be adding a possible version for this modified k-means in the next minutes!

[cesarsouza]

I've just committed a Balanced K-Means implementation. You can either compile the framework from the source code right now or wait until a new pre-release NuGet package is generated.

Regards,
Cesar

[jagbarcelo]

Awesome! Thanks a lot for such a quick response and even quicker solution.

[cesarsouza]

Fixed in 3.5.0.