Implementation in TF2 of Community-Centric Graph Convolutional Network for Unsupervised Community Detection

You can find the paper at the following link https://www.ijcai.org/Proceedings/2020/0486.pdf

To train the model you have to run the following command:

python gcn_clustering.py

by using this command you are using the constants defined in the constants.py file that are not finetuned. I suggest to try to reduce LR and number of parameters.

Other Parameters

• --dataset= to define the dataset (one value between cora and citeseer)
• --lr= to define the learning rate
• --lambda= to define the lambda constant defined in the paper (the variable that defines how to balance Stopo and Satt)
• --gamma= to define the gamma constant defined in the paper (the variable that defines how to balance att loss and topo loss)
• --eta= to define the eta constant defined in the paper (the variable that defines the imporance of the reg loss)
• --beta= to define the beta constant defined in the paper (the variable that defines how to balance topo info and att info in MRF layer)
• --epochs= to define the number of epochs

So you can try to run the following command:

python gcn_clustering.py --dataset="cora" --lr=0.0001

pip3 install scipy numpy tensorflow pandas networkx seaborn sklearn matplotlib munkres