Disconnected communities with ModularityVertexPartition on very large graphs about leidenalg

[szhorvat]

@vtraag, does this affect the implementation in igraph as well? I'm guessing yes?

[vtraag]

Thanks for the report here @wolfram77!

The problem here is that I want to ensure that the diff_move() function is consistent with the difference in the quality() function after the exact same move. Indeed, for the ModularityVertexPartition, the diff_move() is scaled (i.e. with $1/m$ with $m$ the number of edges) in order to be consistent with the quality() function.

What I'll probably do is separate the diff_move function into two separate parts, one diff_move_raw, which can then be used in the optimisation procedure, and one diff_move which is the scaled version of diff_move_raw used for consistency checks.

@vtraag, does this affect the implementation in igraph as well? I'm guessing yes?

No, this does not affect the igraph implementation, it always uses an unscaled version of the difference (i.e. it doesn't divide by $m$).

[wolfram77]

Hello @vtraag I was thinking if it would be easier to recommend users to use RBConfigurationVertexPartition instead of ModularityVertexPartition for very large graphs - until you see enough requests for it to directly work with ModularityVertexPartition. If so, we can update the documentation and close this issue immediately.