Fed-RGCN

Pytorch-based and DGL-based implementation of Relational Graph Convolutional Networks via federated learning for Node Classification

Dependencies

• Python 3.6+
• PyTorch 1.0+
• DGL 0.6+

Results

The first experiment

Dataset Settings

• Ratio of the number of labeled nodes on each client to the total number of labeled nodes: 70%
• Each client's local subgraph does not intersect with other clients

The experimental results

The results of Entire-RGCN, Single-RGCN and the baseline FL Fed-RGCN on AIFB are as follows.

	Test Acc
Entire	0.8611
Single	0.3583
Fed	0.4166

The second experiment

Dataset Settings

• Ratio of the number of labeled nodes on each client to the total number of labeled nodes: 70%
• The ratio of the local subgraph to the full graph for each client: 70%

The experimental results

The results of Entire-RGCN, Single-RGCN and the baseline FL Fed-RGCN on AIFB are as follows.

	Test Acc
Entire	0.9176
Single	0.8444
Fed	0.8833

Conclusion

Entire-RGCN > Fed-RGCN > Single-RGCN

Running the code

Usage

• running Entire-RGCN

cd ./src
python main.py --run_mode=Entire

• running Single-RGCN

cd ./src
python main.py --run_mode=Single

• running Fed-RGCN

cd ./src
python main.py --run_mode=Fed