Hierarchical Clustering to select node clusters in SOM

[Kyle Thayer]

After making an SOM that has many hit histograms with many data points on each, you can save out the hit histogram values for each node. This can then be used to with hierarchical clustering and displayed on a heat map. I found this approach in a paper called "A Microfluidic Platform for Systems Pathology: Multiparameter Single-Cell Signaling Measurements of Clinical Brain Tumor Specimens": http://cancerres.aacrjournals.org/content/70/15/6128.abstract. You can see the one they made in Figure 5.

I spent some time experimenting with it using Flow Cytometry Cell data. I added an option to save the SOM and hit histogram values (http://code.google.com/p/ifcsoft/source/detail?r=92104e7f9d1e5968421618ca253f39bcacc515b9). Since they blurred their hit histogram once before using it, I did as well (that is an option when saving the values). I got the values into R and used the heatmap function to do hierarchical clustering and display the results. I noticed that the node clusters it gave were very good natural selections on the SOM, better than any other automatic node selection technique I've tried.

I think it would be good to include an option to use hierarchical clustering based in the (possibly blurred) hit histogram values for selection of groups of nodes on SOMs.

Kyle