This is a PyTorch implementation of AdaBoost GNN-to-MLP Knowledge Distillation (AdaGMLP) which is built on the source code of KRD (https://github.com/LirongWu/KRD), and the code includes the following modules:

• Dataset Loader (Cora, Citeseer, Pubmed, Amazon-Photo, Coauthor-CS, Coauthor-Phy, and ogbn-arxiv)

• Two evaluation settings: transductive and inductive

• Various teacher GNN architectures (GCN, SAGE, GAT) and student MLPs

• Training paradigm for teacher GNNs and student MLPs

• numpy==1.21.6
• scipy==1.7.3
• torch==1.7.1
• dgl == 0.6.1

• train_and_eval.py

  • train_teacher() -- Pre-train the teacher GNNs
  • train_student() -- Train the student MLPs with the pre-trained teacher GNNs
  • adagmlp_train_mini_batch() -- Train AdaGMLP student model
  • adagmlp_evaluate_mini_batch() -- Evaluate AdaGMLP student model

• models.py

  • AdaGMLP() -- AdaGMLP student
  • MLP() -- student MLPs
  • GCN() -- GCN Classifier, working as teacher GNNs
  • GAT() -- GAT Classifier, working as teacher GNNs
  • GraphSAGE() -- GraphSAGE Classifier, working as teacher GNNs

• dataloader.py

  • load_data() -- Load Cora, Citeseer, Pubmed, Amazon-Photo, Coauthor-CS, Coauthor-Phy, and ogbn-arxiv datasets

• utils.py

  • mask_features() -- Randomly mask a portion of features
  • set_seed() -- Set radom seeds for reproducible results
  • graph_split() -- Split the data for the inductive setting

1. Install the required dependency packages

2. To reproduce the results in paper, please use the command in ./train.sh