Resampling to evaluate clustering stability

[bednarsky]

Summary:
Clustering stability could be assessed by doing multiple clusterings, always randomly sampling 90% of the data.
Consensus approach could be used to extract stable clustering.

Drawbacks:
Computationally expensive

Open questions:
How to combine this with clustification? Should it be only used to evaluate stability of one approach, or automatically to generate consensus clustering?

Background:
Daria implemented a resampling strategy, because she noted that adding two new samples completely changed her previous clustering. She ended up finding gene programs by seeing which genes are stably co-differential in clusters, and then came up with hard clusters by using thresholds to assign cells to one or multiple cluster labels.