Attributed Graph Clustering via Modularity Aided Coarsening

Submitted to TMLR

Instructions to run

This repository consists of 4 main files: q-fgc.py, q-gcn.py, q-vgae.py and q-gmm-vgae.py corresponding to the Q-FGC, Q-GCN, Q-VGAE and Q-GMM-VGAE methods. These are intended to be run directly.

To specify the dataset, set it as an environment variable. For eg. In Linux:

dataset=CiteSeer python q-vgae.py --alpha=1000 --beta=100

And in Windows:

set dataset=CiteSeer
python q-vgae.py --alpha=1000 --beta=100

(equivalently, you can use export in Linux).

For small datasets (Airports), since there are very less number of samples (~100-300), deep learning methods are more sensitive to random state, whereas Q-FGC is not.

Note that all the datasets are sensitive to some extent as the loss function used is non-convex.

Make sure to have PyTorch 1.8+ installed, preferably 2.0+, and PyTorch Geometric 2.0+. Then you can run

pip install -r requirements.txt