search

gephi / gephi-plugins

Repository for Gephi Plugins maintained by the team. Each plugin has it's branch.
270 stars 620 forks source link

Leiden algo: many communities with a single node #203

Closed seinecle closed 5 years ago

seinecle commented 5 years ago

The algo finds many communities which are made of just one node, even when these nodes are not isolated. That is not very useful and actually on large networks, this creates hundreds of communities which make the results impractical to process further.

eduramiba commented 5 years ago

Hi, when testing it, I found that the resolution parameter must be tweaked for every network and is quite sensitive, so you must test a few numbers until you find one that suits your network.

@vtraag can you provide some help?

vtraag commented 5 years ago

Ah, @seinecle, I thought the problem was that it slowed down Gephi?

The single community nodes may simply not fit anywhere else better. Indeed this will be affected by the resolution parameter. Louvain may leave these nodes in larger communities, while they shouldn't.

@eduramiba, is it convenient to keep this issue here? Or is vtraag/gephi-leiden-plugin more suitable for tracking issues around this specific plugin?

eduramiba commented 5 years ago

@vtraag I think your repository would be a better place, yes. Is there a strategy to choose a good resolution parameter?

seinecle commented 5 years ago

Hey! I've written the issue here as I could not see the issues in the menu of vtraag/gephi-leiden-plugin, but I'll check again. There were two issues: lots of communities with jus one node, and slowness of treatment after the report popup. Maybe that it could be left to the user to delete / not create communities that have a member count < n?

(yes, can't see issues in vtraag/gephi-leiden-plugin)

vtraag commented 5 years ago

Sorry, I didn't realize the issue tracker wasn't open! I've now opened it at https://github.com/vtraag/gephi-leiden-plugin/issues. If we move it there, we can close this one, and continue the conversation there. Thanks!

vtraag commented 5 years ago

@eduramiba regarding a strategy good resolution parameter for CPM, this is always a bit trial and error. There are some ways to scan entire resolution ranges, to see how things change for different resolution parameters (I've written about that here). As a starting point, a resolution parameter around the average edge weight (i.e. graph density), should be reasonable. But there is no magic formula unfortunately, and any resolution parameter may potentially be of interest.

I should add that for the Modularity quality function the default resolution parameter of 1 is also quite arbitrary in a sense, even though that may similarly be a good starting point for further exploration.

vtraag commented 5 years ago

Moved to https://github.com/vtraag/gephi-leiden-plugin/issues/2, @eduramiba, you can close this one.