K-Centroids Clustering - Number of starting seeds

[ashi_s]

Hi, 

I am actually trying to create a few clusters with my data of about 400 observations, and am a little confused about the role played by the starting number of seeds. 

Irrespective of changing my number of starting seeds I end up with the exact same observations in each cluster. Which is why I was trying to understand if starting the analysis with a different value for the number of seeds can actually change my final clusters. And if yes, when does it happen ? 

Thanks

[jdunkerley79 · Accepted solution]

The K-Means algorithm can converge to local minimums depedning on the input points and the starting seed. Having different starting seeds and then picking the `best` K-mean solution is the standard way to work around this. For you specific dataset it might be that it always converges to the same global minimum.

A more detailed post: https://stats.stackexchange.com/a/50316