Assessing cluster stability/robustness

Any suggestions on methods that might help in accessing clustering stability/robustness ?


People seem to like the silhouette score… but that assumes that euclidean distances in some space can tell you sth about cluster stability as well. There is also a recent approach from the Teichmann lab based on building classifiers for a cluster from marker genes. The performance of the classifier is an assessment of the cluster stability.

1 Like

Thanks, sorry I somehow missed your response. I’ll give this a try.

If you’re still looking at this, I should probably self promote a bit here :blush:

The idea here is that robust clusters should keep showing up if you change the parameters a little. Turns out a lot clusters do not do this. Some samples of a dataset may cluster well with a particular parameter, while you’ll get junk results for others.

Congrats @ivirshup ! Do you also do a consensus over random seeds in this?


Yep, those are included as “unordered parameters”. I think it’s the only kind of unordered parameter it makes sense to use though.