I am looking to do community detection to identify clusters and themes in technical papers. I have started with Louvain endnote real happy with the stability of the results.
What I did was to run the algorithm multiple times and look at the size of the top communities. It the graph I plot the community size as the y variable and the rank of the community (by size) as the x variable. You see quite a bit of spread in size of the communities especially the large ones which are the ones that matter. The simplistic fix is just to fix the seed parameter, however that just pushes the problem to a different spot which seed is "correct". What is a more robust algorithm to use?
I did a quick look at label propagation and it appears on first blush more stable using number of nodes in community as a metric.