Spectral clustering: K-means replicates setting to improve reliability

Hello,

The command spectralcluster in matlab can use k-means as part of its algortihm (and is the default in the options ClusterMethod — Clustering method). In the documentation for k-means there is a way to pass 'Replicates' as an option to the algorithm to avoid problems arising from a bad centroid initialization. Any way to pass this replicates option to spectral clustering so it is not only reproduceable but also more reliable? It would be way better than just specifying the rng seed but I cannot figure out the correct syntax. Should it be done with an options struct?

Thanks in advance!

Answer 1

[Adam Danz]

If you open spectralcluster, you'll see that kmeans is being called with a Replicates value of 5 and there is not a way to specify a different replicates value in spectralcluster. Kmeans selects a random starting point which can be controlled by the Options>Streams option in kmeans() but this cannot be specified as inputs to spectralcluster.

Note the examples in the spectralcluster doc page uses rng() to control reproducibility.

[Gustavo Lunardon]

Thanks for the explanation!