Louvain algorithm

[jalving]

Here is a simple implementation of the Louvain fast-unfolding algorithm I whipped up. It has been giving valid communities for my problems.

The algorithm could use some speed-up enhancements. Notably, I'm using the LightGraphs.modularity function to calculate modularity change in my loop as opposed to calculating the change directly.

Please feel free to comment or make changes.

[simonschoelly]

I think your pr kind of went unnoticed, as there is not much activity on this package at the moment. I will try to have a look at your pr soon.

[jalving]

No rush, I've been using community detection methods in my research and figured this is the best place to share the work.

[VPetukhov]

Hi @simonschoelly , do you by chance have any updates on this?

[storopoli]

I did some benchmarking and the algorithm is performing very well

using BenchmarkTools
G = erdos_renyi(100, 0.1)

@benchmark community_detection_louvain(G)
@benchmark community_detection_bethe(G)

Benchmark for Bethe

BenchmarkTools.Trial: 
  memory estimate:  523.06 KiB
  allocs estimate:  266
  --------------
  minimum time:     1.485 ms (0.00% GC)
  median time:      1.766 ms (0.00% GC)
  mean time:        1.943 ms (3.15% GC)
  maximum time:     46.315 ms (0.00% GC)
  --------------
  samples:          2572
  evals/sample:     1

Benchmark for Louvain

BenchmarkTools.Trial: 
  memory estimate:  181.54 MiB
  allocs estimate:  77601
  --------------
  minimum time:     337.115 ms (7.82% GC)
  median time:      352.989 ms (7.50% GC)
  mean time:        355.509 ms (7.61% GC)
  maximum time:     382.444 ms (5.86% GC)
  --------------
  samples:          15
  evals/sample:     1

I think it's time to do a pull request and make it to the codebase

[storopoli]

Is it possible to use a memoing function or to have the current_modularity cached? I think this is the major bottleneck in performance. Did you profiled the code to inspect what are the major bottlenecks?

[jalving]

Thanks for taking a look at the algorithm. I never did any profiling but it shouldn't be too bad to check the bottlenecks. I'll take a look at improving the benchmark. Maybe we can merge once we're happy with the performance?

[storopoli]

Try Memoize.jl

[gdalle]

Closed in favor of #4