1
$\begingroup$

The main goal is to cluster subjects based on several distinct cross-correlation matrices.

So I have 4 correlation-matrices, each corresponding to dissimilarity $(1-r)$ of brain topography between subjects, (i.e. $s\times s$ cross-correlation matrix, $s =$ subjects), during a specific task (4) (i.e. $s\times s$ matrices for 4 tasks)

My first instinct would've been to perform hierarchical clustering on each matrix and then perform another clustering method on the resulting cluster memberships. However, i do feel that there's probably something better that can be done here.

I have around 1,000 subjects (so 1,000 subjects by 1,000 subjects matrix for each of the 4 tasks)

I could convert the 4 tasks in a 1D feature vector and then correlate between subjects but in doing so, i will lose a lot of information.

Thank you very much

$\endgroup$
1
$\begingroup$

Try interclustering the de-noised or de-toned correlation matrix (see random matrix theory) and then intraclustering those clusters. In other words, order variables into groups along the diagonal of the matrix (interclustering), and then order the variables within those groups (intraclustering). this significantly stabilizes the matrix compared to hierarchical clustering alone because it alleviates signal-induced instability as opposed to only noise-induced instability

$\endgroup$
1
  • $\begingroup$ Thanks for this suggestion! However i'm still stuck on which model could apply. Since clustering models are used on a single matrix, do you have any suggestions on how to do this (clustering 4 matrices)? $\endgroup$ – Jules D. Sep 18 '20 at 15:14

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.