Pytorch-based and DGL-based implementation of Relational Graph Convolutional Networks via federated learning for Node Classification
- 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 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 |
- 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 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 |
Entire-RGCN > Fed-RGCN > Single-RGCN
- 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