Score to Evaluate Clustering Stability (rather than tightness), as Summary of Co-Clustering Matrix

[drisso]

Score to Evaluate Clustering Stability (rather than tightness), as Summary of Co-Clustering Matrix

[mbcole]

Motivation for entropy measure is to distinguish block-diagonal co-clustering matrices in a clustering-stability analysis. Entropy-based methods came up in a discussion with Davide.

Instead of entropy, implemented an evaluation of the "compactness" of the zero-region from the (sorted) co-clustering matrix, using an approximate isoperimetric quotient (PAM_COMPACT). Maximum compactness criterion pushes the one-region (co-clustering matrix elements) toward the diagonal. Indeed, this score rewards methods that overfit for biological condition, as does the previously-implemented PAM_SIL score. On the other hand, it may still have better properties than PAM_SIL.

To be determined...