This is an official implementation of Multi-hop Attention Graph Neural Network (IJCAI2021).

Implementation Environment:

• python >= 3.7
• torch >= 1.4.0
• numpy == 1.17.2
• dgl-cu101 == 0.4.3

Tasks: Node classification and KG completion

• Node classification

•  datasets: Cora, Citeseer, Pubmed (from DGL dataset)
  

•  codes --> node classification (train.py)
  

• Knowledge Graph Completion

•  datasets: Freebase237-15k and WordNet18RR (from https://github.com/villmow/datasets_knowledge_embedding)
  

•  codes_kge --> KG emebdding (all the pakages with 'kge' are used for Knowledge graph embedding) (runkge.py)
  

• parametertuning --> random search based hyper-parameter tuning

Hyper-parameter demonstration for MAGNA

• in_feats: dimension of input node features (for KG embedding: in_ent_feats, in_rel_feats represent the dimension of
• entity embedding and relation embedding, respectively)
• hidden_dim: dimension of hidden dimension
• num_heads: head number of multi-head attention,
• alpha: transition probability in personal page-rank (0.05 ~ 0.25, data dependent)
• hop_num: number of iterations to approximate (3 ~ 10, data dependent)
• feat_drop: dropout ratio over the feature (input node features)
• attn_drop: dropout ratio over attention matrix
• topk_type='local': if the degree as nodes is large. ('local': select top-k for each head, otherwise select top-k
• shared by all heads)
• top_k=-1: top-k neighbor selection, -1 means no 'top-k' selection is performed

References

@inproceedings{wang2020multi, title={Multi-hop Attention Graph Neural Network}, author={Wang, Guangtao and Ying, Zhitao and Huang, Jing and Leskovec, Jure}, booktitle={International Joint Conference on Artificial Intelligence}, year={2021} }

Related resources

• GAT in DGL: https://docs.dgl.ai/en/0.4.x/tutorials/models/1_gnn/9_gat.html
• APPNP: https://github.com/klicperajo/ppnp, https://github.com/dmlc/dgl/tree/master/examples/pytorch/appnp (DGL implementation)
• RotatE: https://github.com/DeepGraphLearning/KnowledgeGraphEmbedding