Can we use k-means clustering on a 100×34596 matrix?

• I have a feature matrix with size 100×34596, with class labels.
• This matrix is very large because I want to cluster these matrix.
• Is it possible?
• I don't have experience about k-means.
• Could you help me?*

Answer 1

[Walter Roberson]

Yes. See https://www.mathworks.com/help/stats/kmeans.html

Just make sure to pass data to kmeans such that the rows correspond to observations (points).

The class table will not be of any use for kmeans clustering: kmeans is a method of trying to figure out which values belong together without knowledge of the class labels.

[Image Analyst]

Hopefully you have 34,596 observations and 100 features (which still seems ridiculously high) and not 100 observations of 34,596 features (which I don't see as possible in most normal/typical scenarios).

Anyway, if you have some of them labeled already, then you can use knnsearch() to label/classify any other, non-labeled data.

kmeans will come up with its own labels, so if they're all labeled already, then all you can do is pass in the feature values, without the labels as Walter said, and see how what kmeans chose differs from whatever method you have that produced the labels in advance (like human judging or whatever). You can then create a confusion matrix or ROC curve.