12
$\begingroup$

I've had a hard time interpreting resulting clusters of an adjacency matrix. I have 200 relatively big matrices representing subjects that contains partial correlations (z scores) of time series (neural data). The goal is to cluster those 210 matrices and detect any potential undiscovered communities. So I did another partial correlation calculations resulting in 200x200 adjacency matrix. Whenever I run a community detection algorithm (eg Newmann's) it comes up with hardly interpretable communities.

The question is that what kind of statistical tests that will tell if these communities or clusters are significant at all ? and if so, are there systematic ways to work out the interpretation ?

$\endgroup$
3
  • $\begingroup$ As far as I am aware there is no single 'right way' to do this. An approach would be to use something like hierarchical clustering on the distance matrix $1 - |\rho|$ where $\rho$ is the correlations. The other thing is whether your latter correlation matrix will capture meaningful relationships. What steps were taken to produce it? $\endgroup$ Nov 24, 2014 at 17:04
  • $\begingroup$ Thanks. Regarding your question, the thing I did is I correlated every row (or subject's data) with every other subject using corrcoef (simple correlation) and that's how I got the results. I 'm trying to detect the patterns and that's why I had to correlate again. $\endgroup$
    – Fahd
    Nov 24, 2014 at 17:46
  • $\begingroup$ in the OP it is suggested that the subject data is matrix valued so how does this become a single row per subject? $\endgroup$ Nov 24, 2014 at 18:02

3 Answers 3

3
$\begingroup$

I have done some work in the past on spectral clustering which might be of use here. The basic idea is that one can use the adjacency matrix to form the so called Laplacian matrix:

$L = I-D^{-1/2}AD^{-1/2}$

You can check for yourself that the lowest eigenvalue of the Laplacian is zero. The first nonzero eigenvalue is often called the algebraic connectivity, and the corresponding eignevector will have positive part and negative part corresponding to two partitions $(B_1,B_2)$ of the underlying graph. Roughly speaking, the magnitude of the first nonzero eignevalue is a measure of the strength of the connections between the two partitions. Perhaps you could employ this approach recursively or consider the first few nonzero eignvalues of the Laplacian.The following Wikipedia article about spectral clustering is a good start.

$\endgroup$
2
$\begingroup$

I am looking at the same problem at the moment. From quick review, it seems like Spectral clustering is the most "natural" way to analyze an Adjacency matrix. See this blog post for more details.

$\endgroup$
0
$\begingroup$

Alternatively... Neural data (real or artificial) is often a highly compressed representation of data, which means the data is very random, which means you won't find any correlations. Which you have!! Congratulations! :)

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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