Community with Edge Weight

[elijah]

Any chance that there's going to be a Modularity analytic that takes edge weight into account? As I understand it, Louvain does not, and this bears out to my experience with very (very) low-weight edges significantly reducing statistical significance of identified modules.

[mbastian]

That would be useful indeed,

Currently Louvain method in Gephi doesn't use edge's weight

[elijah]

It would be incredibly useful. For very complex networks, I lean heavily on modularity, and it's lack of taking into account edge weight makes a mess of things. I just finished building a 120,000 node, 173,000 edge network focused on 50,000 documents, 10,000 words and 50 topics, and I'm dreading the actual representation of analysis of it with the currently modularity metric (edge weight ranges from .08 to 1.0). I've found workarounds with centrality and degree range, but mostly I end up dropping a broad swathe of low-weight edges to build the modules first.

To be even more ambitious, I'd like to see something that takes edge hierarchies into account as well. I categorize and supercategorize my edges on certain semantic graphs and if the modularity could somehow take this into account (rather than, say forcing the user to come up with some kind of arbitrary numerical identifier for edge significance when such an identifier is a crude representation of a semantic value) then I'd be extremely excited.

Which analytics and layouts take edge weight into consideration? I just realized I've been assuming when it's really unclear to me (I know Yifan Hu doesn't and ForceAtlas does).

[seinecle]

+1 !

[admin]

The official Louvain algorithm takes edges weights into account to build a hierarchy of modules in N levels.

So I guess we might fix that in our current implementation. I filled the Ideas List

[elijah]

That would be incredibly valuable for my work. As it stands, I'm forced to pre-process my own hierarchy based on natural breaks in the edge weight, and it is tedious work.

[incunix]

Any news on this?

Cheer,
Nick

[incunix]

My bad - already implemented! Good work!
https://github.com/gephi/gephi/issues/556

Nick