Comparison experiment

[haory-95]

Hi, I recently read your paper,I‘m very interested in your paper,Can you provide comparison experiment code?For example,I I reproduced the result of Bigclam method is very low（ NMI=0.05.But the result you get is NMI=0.26）So I am very curious about what I did wrong.Thank you very much

[shchur]

Hi, the link to the TensorFlow 1.0 code used to run the experiments is provided in README.md (https://figshare.com/s/30894e4172505d5dc070). I haven't ported the baselines like BigCLAM to Pytorch, but that should be relatively easy - just make F a learnable nn.Parameter and clip the negative values to zero after each gradient descent step.

I don't remember the details very well now – I wrote it about 3 years ago – but I think the two main reasons for poor performance of BigCLAM can be:

1. Poor initialization. You can have a look at the original code to see how the F matrix was initialized there for BigCLAM.
2. Poor choice of the threshold for assigning nodes to communities after training. IIRC, 0.5 was a good choice for balanced loss (edges balanced with non-edges), otherwise it was either 0.01 or 0.1, I don't remember exactly.