*3.7. Training Algorithm*

Algorithm 1 illustrates the proposed framework. **SORAG** is trained through the following components: (1) the selection of seed examples based on node LSP and USP scores; (2) the pre-training of the node generator (i.e., the ensemble of GAN and CGAN) for synthetic data generation; (3) the pre-training of the edge generator to produce new relation information; and finally, (4) the training of the node classifier on top of the over-sampled graph and the fine-tuning of the node generator and edge generator. The computational complexity of our model is approximately the sum of the computational complexity of the contained GAN, CGAN, and GCN.
