Community Detection - Revision history

Chris: /* Minimum Connected Component */

2020-11-05T22:42:40Z

Minimum Connected Component

← Older revision		Revision as of 18:42, 5 November 2020
Line 65:		Line 65:

	====Minimum Connected Component====		====Minimum Connected Component====
	A Minimum Connected Component is a subset of your weight matrix that retains the largest weights in descending order until every vertex is connected to at least one other vertex (at which point we stop adding weights). Algorithmically, this might take a while, if you go by the letter of the algorithm, but it turns out there's a trivial shortcut. Note that this code will not work if your connectivity matrix only includes the upper or lower-triangle.		A Minimum Connected Component is a subset of your weight matrix that retains the largest weights in descending order until every vertex is connected to at least one other vertex (at which point we stop adding weights). Algorithmically, this might take a while, if you go by the letter of the algorithm, but it turns out there's a trivial shortcut. Note that this code will not work properly if your connectivity matrix only includes the upper or lower-triangle.
	mcc=abs(weights); %absval because we care about magnitude		mcc=abs(weights); %absval because we care about magnitude
	mcc(eye(size(mcc,1))==1)=0; %remove self-connections		mcc(eye(size(mcc,1))==1)=0; %remove self-connections

Chris: /* Minimum Connected Component */

2020-11-05T22:42:19Z

Minimum Connected Component

← Older revision		Revision as of 18:42, 5 November 2020
Line 65:		Line 65:

	====Minimum Connected Component====		====Minimum Connected Component====
	A Minimum Connected Component is a subset of your weight matrix that retains the largest weights in descending order until every vertex is connected to at least one other vertex (at which point we stop adding weights). Algorithmically, this might take a while, if you go by the letter of the algorithm, but it turns out there's a trivial shortcut.		A Minimum Connected Component is a subset of your weight matrix that retains the largest weights in descending order until every vertex is connected to at least one other vertex (at which point we stop adding weights). Algorithmically, this might take a while, if you go by the letter of the algorithm, but it turns out there's a trivial shortcut. Note that this code will not work if your connectivity matrix only includes the upper or lower-triangle.
	mcc=abs(weights); %absval because we care about magnitude		mcc=abs(weights); %absval because we care about magnitude
	mcc(eye(size(mcc,1))==1)=0; %remove self-connections		mcc(eye(size(mcc,1))==1)=0; %remove self-connections

Chris: /* Minimum Spanning Tree */

2018-10-26T17:25:00Z

Minimum Spanning Tree

← Older revision		Revision as of 13:25, 26 October 2018
Line 40:		Line 40:
	This was an old method used in one of the Sporns group's papers. If I recall correctly, the logic was something like: "10% of all possible brain connections exist, so we will retain the top 10%-ile of the weights (by magnitude), and set all other weights to zero".		This was an old method used in one of the Sporns group's papers. If I recall correctly, the logic was something like: "10% of all possible brain connections exist, so we will retain the top 10%-ile of the weights (by magnitude), and set all other weights to zero".

	====Minimum Spanning Tree====		====Minimum Spanning Tree (MST)====
	A graph-theoretic-based approach, this starts with an empty weight matrix, and adds weights according to an algorithm that I'll have to look up. MATLAB has a minspantree function requiring a Graph. You can create a graph from a vector, '''S''', of source nodes, '''T''' of target nodes, and a bidirectional (i.e., nonredunant) set of weights:		A graph-theoretic-based approach, this starts with an empty weight matrix, and adds weights according to an algorithm that I'll have to look up. MATLAB has a minspantree function requiring a Graph. You can create a graph from a vector, '''S''', of source nodes, '''T''' of target nodes, and a bidirectional (i.e., nonredunant) set of weights:
	S=[1 1 1 2 2 3];		S=[1 1 1 2 2 3];
Line 48:		Line 48:
	[T,pred]=minspantree(G);		[T,pred]=minspantree(G);

	~~Importantly~~, the weights in the graph represent distance. Distance is generally inversely proportional to strength, so if you have a weight matrix, '''W''', where big numbers represent strong connections, you probably want to create your graph using the inverse of your weights:		'''NOTE:''' minimum spanning trees contain no cycles, by definition. This is going to be quite relevant for some of your network metrics (e.g., those that count triangles, which form a cycle), so consider this before using MST in your network pruning.

			Another Important thing to consider is that the weights in the MST graph represent distance. Distance is generally inversely proportional to strength, so if you have a weight matrix, '''W''', where big numbers represent strong connections, you probably want to create your graph using the inverse of your weights:
	G=graph(S,T, power(weights,-1);		G=graph(S,T, power(weights,-1);

Line 56:		Line 58:
	ADJ=corr(timeseriesdata);		ADJ=corr(timeseriesdata);
	%produce the minimum spanning tree for this matrix		%produce the minimum spanning tree for this matrix
			ADJ=power(ADJ,-1); %calculate the reciprocal of the connection strengths so that strong weights are mapped to short distances
			%You may also need to eliminate or compute the absolute value of negative connections
	[mst, G, T, pred]=minspan(ADJ, 'method','sparse','root',1, 'type', 'forest');		[mst, G, T, pred]=minspan(ADJ, 'method','sparse','root',1, 'type', 'forest');

Chris: /* Minimum Spanning Tree */

2018-10-26T17:17:34Z

Minimum Spanning Tree

← Older revision		Revision as of 13:17, 26 October 2018
Line 56:		Line 56:
	ADJ=corr(timeseriesdata);		ADJ=corr(timeseriesdata);
	%produce the minimum spanning tree for this matrix		%produce the minimum spanning tree for this matrix
	[mst, G, T, pred]=minspan(ADJ);		[mst, G, T, pred]=minspan(ADJ, 'method','sparse','root',1, 'type', 'forest');

			The only required parameter is '''ADJ'''. The default minimum spanning tree produced uses the same parameters specified in the example above: it will be sparse, with the root node set to node 1, and returning a forest (i.e., multiple trees, if that's how the weights shake out). See the help for the MATLAB minspantree() function for details on what these options mean.

	====Minimum Connected Component====		====Minimum Connected Component====

Chris: /* Creating .node Files */

2018-09-13T23:59:31Z

Creating .node Files

← Older revision		Revision as of 19:59, 13 September 2018
Line 201:		Line 201:
	%this gives the indices of the elements in lausanne_labels		%this gives the indices of the elements in lausanne_labels
	%that overlap with those in cluster_labels		%that overlap with those in cluster_labels
			%You can inspect for yourself by calling:
			%(labels(r))

	M_reduced=M(r,r); %this selects the subset of M corresponding to the overlapping node indices		M_reduced=M(r,r); %this selects the subset of M corresponding to the overlapping node indices
	%note that they will now also be sorted in the same order as they appear in '''r'''		%note that they will now also be sorted in the same order as they appear in '''r'''

Chris: /* Creating .node Files */

2018-09-13T20:43:56Z

Creating .node Files

← Older revision		Revision as of 16:43, 13 September 2018
Line 203:		Line 203:
	M_reduced=M(r,r); %this selects the subset of M corresponding to the overlapping node indices		M_reduced=M(r,r); %this selects the subset of M corresponding to the overlapping node indices
	%note that they will now also be sorted in the same order as they appear in '''r'''		%note that they will now also be sorted in the same order as they appear in '''r'''
	%This means that the first region in '''M_reduced''' will probably not be the same as the first region in '''cluster_labels'''		%This means that the first region in '''M_reduced''' will probably not be the same
			%as the first region in '''cluster_labels'''

	If we analyzed communities within '''M_reduced''' in the context of the Lausanne parcellation, then the relevant node information would be in XYZ(r), and labels(r). Here's a snippet of code that will generate a .node file:		If we analyzed communities within '''M_reduced''' in the context of the Lausanne parcellation, then the relevant node information would be in XYZ(r), and labels(r). Here's a snippet of code that will generate a .node file:

Chris: /* Creating .node Files */

2018-09-13T20:43:39Z

Creating .node Files

← Older revision		Revision as of 16:43, 13 September 2018
Line 198:		Line 198:

	This is documented elsewhere, but repeated here: If you have two sets of cell arrays (e.g., '''lausanne_labels''' and '''cluster_labels'''), and you want to find out which elements in lausanne_labels map to the elements in cluster_labels, use the <code>ismember</code> MATLAB function:		This is documented elsewhere, but repeated here: If you have two sets of cell arrays (e.g., '''lausanne_labels''' and '''cluster_labels'''), and you want to find out which elements in lausanne_labels map to the elements in cluster_labels, use the <code>ismember</code> MATLAB function:
	r=find(ismember(lausanne_labels, cluster_labels));		'''r'''=find(ismember(lausanne_labels, cluster_labels));
	%this gives the indices of the elements in lausanne_labels		%this gives the indices of the elements in lausanne_labels
	%that overlap with those in cluster_labels		%that overlap with those in cluster_labels
	M_reduced=M(r,r); %this selects the subset of M corresponding to the overlapping node indices'		M_reduced=M(r,r); %this selects the subset of M corresponding to the overlapping node indices
			%note that they will now also be sorted in the same order as they appear in '''r'''
			%This means that the first region in '''M_reduced''' will probably not be the same as the first region in '''cluster_labels'''

	If we analyzed communities within '''M_reduced''' in the context of the Lausanne parcellation, then the relevant node information would be in XYZ(r), and labels(r). Here's a snippet of code that will generate a .node file:		If we analyzed communities within '''M_reduced''' in the context of the Lausanne parcellation, then the relevant node information would be in XYZ(r), and labels(r). Here's a snippet of code that will generate a .node file:

Chris: /* Creating .node Files */

2018-09-13T20:41:06Z

Creating .node Files

← Older revision		Revision as of 16:41, 13 September 2018
Line 193:		Line 193:
	This option is getting promoted because it is centered around the concept of the brain as a network, which is reflected in how the data is visualized. Your first task here is going to be to generate .node (and possibly .edge) files.		This option is getting promoted because it is centered around the concept of the brain as a network, which is reflected in how the data is visualized. Your first task here is going to be to generate .node (and possibly .edge) files.
	=====Creating .node Files=====		=====Creating .node Files=====
	If you haven't already, you will need the XYZ coordinates for each of your labels. If you're working with the 1000 node Lausanne parcellation, I've already computed these and saved them in ''regionlabels.mat'', which you can find in the ubfs/Scripts/Matlab folder. You will also need a way to map each node in your adjacency matrix on to the associated label. If you're working with the full adjacency matrix, then the mapping is trivial: node 1 in your adjacency matrix corresponds to label 1 in the regionlabels matrix. If, however, you've simplified your network (e.g., dropped regions that didn't show significant task activation), ~~then you~~ will need to figure out for example that node 1 in your adjacency matrix is <code>lh.superiorfrontal_27</code>, which is actually entry 398 in the lausanne labels, because that's the only way you're going to figure out the correct X, Y, Z coordinates for the nodes in your reduced network.		If you haven't already, you will need the XYZ coordinates for each of your labels. If you're working with the 1000 node Lausanne parcellation, I've already computed these and saved them in ''regionlabels.mat'', which you can find in the ubfs/Scripts/Matlab folder. You will also need a way to map each node in your adjacency matrix on to the associated label. If you're working with the full adjacency matrix, then the mapping is trivial: node 1 in your adjacency matrix corresponds to label 1 in the regionlabels matrix.

			If, however, you've simplified your network (e.g., dropped regions that didn't show significant task activation), things get a little more complicated. You will need to figure out for example that node 1 in your adjacency matrix is <code>lh.superiorfrontal_27</code>, which is actually entry 398 in the lausanne labels, because that's the only way you're going to figure out the correct X, Y, Z coordinates for the nodes in your reduced network.

	This is documented elsewhere, but repeated here: If you have two sets of cell arrays (e.g., '''lausanne_labels''' and '''cluster_labels'''), and you want to find out which elements in lausanne_labels map to the elements in cluster_labels, use the <code>ismember</code> MATLAB function:		This is documented elsewhere, but repeated here: If you have two sets of cell arrays (e.g., '''lausanne_labels''' and '''cluster_labels'''), and you want to find out which elements in lausanne_labels map to the elements in cluster_labels, use the <code>ismember</code> MATLAB function:

Chris: /* Creating .edge Files */

2018-09-13T20:40:02Z

Creating .edge Files

← Older revision		Revision as of 16:40, 13 September 2018
Line 219:		Line 219:
	=====Creating .edge Files=====		=====Creating .edge Files=====
	This is trivially easy, since BrainNet Viewer .edge files are just tab-delimited text files describing the adjacency matrix:		This is trivially easy, since BrainNet Viewer .edge files are just tab-delimited text files describing the adjacency matrix:
	dlmwrite('ReducedNet.edge', '''~~ADJ~~''', 'delimiter', '\t', 'precision', 3);		dlmwrite('ReducedNet.edge', '''M_reduced''', 'delimiter', '\t', 'precision', 3);

	==== Rendering in FreeSurfer ====		==== Rendering in FreeSurfer ====

Chris: /* Creating .edge Files */

2018-09-13T20:39:37Z

Creating .edge Files

← Older revision		Revision as of 16:39, 13 September 2018
Line 219:		Line 219:
	=====Creating .edge Files=====		=====Creating .edge Files=====
	This is trivially easy, since BrainNet Viewer .edge files are just tab-delimited text files describing the adjacency matrix:		This is trivially easy, since BrainNet Viewer .edge files are just tab-delimited text files describing the adjacency matrix:
	dlmwrite('ReducedNet.edge', ~~normmcc~~, 'delimiter', '\t', 'precision', 3);		dlmwrite('ReducedNet.edge', '''ADJ''', 'delimiter', '\t', 'precision', 3);

	==== Rendering in FreeSurfer ====		==== Rendering in FreeSurfer ====