How can I replicate R’s nstart (multiple random starts) with KNIME’s K-Means node

[Damin]

Hi everyone,

I’m trying to reproduce in KNIME exactly what I do in R with:

km_res <- kmeans(data_scaled, centers = 3, nstart = 1000, iter.max = 200)

That is, run K-Means with 1000 random starts (nstart = 1000), pick the best solution (minimum total within-cluster sum of squares), and then use that model.

So far I’ve attempted:

KNIME’s built-in K-Means node
I can set k = 3, max iterations = 200, and a fixed seed, but there is no nstart or “number of restarts” parameter, nor any checkbox to output the total within-cluster sum of squares for each run.

Counting Loop Start → K-Means → Variable Loop End
I tried mapping the loop index (iteration) as the seed, and hoped to collect the WSS via the K-Means node’s flow variables—but there isn’t actually an option in the K-Means node to append totalWithinSS (or similar) into flow variables.
Without that, I can’t sort/filter to pick the best run automatically.

Question:
Is there a more direct or built-in way to perform multiple random starts with KNIME’s native K-Means (or a common extension), analogous to R’s nstart? (without using Python/R snippets)

Thanks in advance for any pointers!

[Fabien_Couprie]

I would have done something like that :

double click on the component for parameters
kmean.knwf (103.5 KB)

[Damin]

Thanks for your help!
I will try that out.

[Fabien_Couprie]

This way you can have a more complete version with the number of iteration max for each try and column selection for the model.
kmean.knwf (101.1 KB)