Network Analyses

Network analyses use graph-theoretic approaches to identifying/quantifying/describing/classifying networks of any sort. For example, a network analysis could be used to identify potential bottlenecks in a city design when large numbers of people have to evacuate, or the robustness of a distributed computer network to disruption in the event that one or more redundant servers goes offline.

In our lab, we apply these analytic approaches to networks that describe functional (primarily) or anatomical (potentially) connectivity. Under certain assumptions about what functional role various network nodes play, we can test hypotheses about how the connections between these nodes relates to cognitive processing. Most of our work here will be done in MATLAB and using networks that are represented as adjacency matrices. An adjacency matrix is an NxN square matrix that has entries that indicate whether there is a connection (and potentially how strong it is) between any two of the N nodes in the network. Using the Louvain 2008 parcellation scheme, for example, we break the cortical surface up into 1000 regions. Each region is a node in the network, and so the adjacency matrix for this network is a 1000 x 1000 square matrix. What follows assumes that you have obtained an adjacency matrix, one way or another, and wish to carry out different types of network analyses on it.

Hierarchical Clustering

Sarah Muldoon recommended starting off with a hierarchical clustering of the adjacency matrix in order to visualize the structure of the matrix and get a sense of the sort of structure that exists within the data. The MATLAB help page for hierarchical clustering has some background information, but here are the highlights.

The first thing to note is that these steps are too computationally intensive for the adjacency matrices generated for the larger networks produced using the Lausanne 2008 parcellation (Lausanne 500 and 250 are too large). The myaparc_60 network seems to work smoothly; I'll try the 125 next.

Obtain Linkage

The first step of a hierarchical cluster is to obtain some kind of distance metric between each node in the graph. This information already exists in the form of the weighted adjacency matrices (either correlations or trained network weights). We pick up here at the linkage step. MATLAB has a linkage function that expects a vector of pairwise distances between all nodes, as would be generated by the pdist function. If you read the documentation for pdist, you will find that it outputs a vector for all pairs ({2,1}, {3,1}, ... , {m, 1}, {3, 2}, ..., {m, 2}, ... , {m, m-1}), where m is the number of nodes in the network. In relation to the adjacency matrices, these values would represent the lower (or upper) triangle of the matrix.

Filter Weights

The lower triangle of the adjacency matrix can be obtained in MATLAB using the tril function. An mxm (i.e., square) matrix has k diagonals, numbered 0 to (-1)*m-1. The main diagonal, k=0, will be the diagonal containing all ones (in a correlation matrix), which are meaningless in this context. We will want to get all the diagonals where k<0 (if you are getting the upper triangle using the triu function, use k>0).

mask=tril(ones(size(M)),-1); %identify the elements in the lower triangle, excluding diagonal k=0
lowert=M(find(mask)); %use mask to obtain the values.

The above code returns the values from the lower triangle of the adjacency matrix M as a vector in the same order that pdist generates.

Run linkage

There are several methods for running the linkage function, but common choices are single, complete, and average.

Z=linkage(lowert,'average');

Note that the solutions provided by each of these methods can yield drastically different results. For example, while I was working through this workflow for the first time, I used the single linkage distance calculation method. The sorted correlation matrix displayed at the end didn't seem to have any well-defined clusters of any size (leading me to assume that I had done something wrong). Since there doesn't seem to be any "correct" method, and the purpose of this exercise is to identify some sort of underlying structure to the weights, you might try different methods until you find something that makes that structure apparent.

Generate the Dendrogram

Using the dendogram function, you'll need to force it to provide the entire leaf structure, specifying p, the maximum number of leaf nodes:

[H, T, outperm] = dendrogram(Z, size(M,1));

Sort and Visualize Adjacency Matrices

Use the outperm vector generated by the dendrogram as a sorting index for the adjacency matrix to reveal the network clustering structure.

sorted_M=M(outperm,outperm); 
figure(2);
imagesc(sorted_M);