This is the pytorch implementation of Dynamic Transfer Learning across Graphs
Please download the datasets from the original paper listed in our paper. And put them under ''./data'' folder
DyTrans is firstly pre-trained on the source dataset for 2000 epochs; then it is fine-tuned on the target dataset for 600 epochs using limited labeled data in each class. We use Adam optimizer with learning rate 3e-3. AUC is used as the evaluation metric.
python dytrans.py --datasets='D3+D5' --finetune_epoch=600 --mu=1e-2 --gnn='gcn' --few_shot=5 --epoch=2000 --heads=4 --m_dim=128 --feat_num=128 --batch_size=-1 --finetune_lr=0.01 --ratio 0.7 --_alpha=0.01 --_alpha=0.01 --only True
python dytrans.py --datasets='Br+D3' --finetune_epoch=600 --mu=1e-2 --gnn='gcn' --few_shot=5 --epoch=2000 --heads=4 --m_dim=64 --feat_num=128 --batch_size=-1 --finetune_lr=0.01 --ratio 0.7 --_alpha=0.05 --_alpha=0.05 --only True