search

Shen-Lab / GraphCL

[NeurIPS 2020] "Graph Contrastive Learning with Augmentations" by Yuning You, Tianlong Chen, Yongduo Sui, Ting Chen, Zhangyang Wang, Yang Shen
MIT License
547 stars 103 forks source link

efficience about data augmentation on Large graph #3

Closed junkangwu closed 3 years ago

junkangwu commented 3 years ago

Hi, I'm interested in self-supervised learning on the graph and GraphCL is an excellent work that combining self-supervised learning with augmentation. However, I'm also curious about its usage in large graphs such as knowledge graph and so on. May I ask you about the suggestion on the augmentation operation on the knowledge graph? Because in the way of GraphCL, it needs 4 GNN encoders at the same time per epoch. Thanks a lot in advance~

yyou1996 commented 3 years ago

Hi @Wjk666,

Thanks for your interest. For NodeDrop, EdgePert and AttrMask augmentations, complexity should be O(K) (where K is the augmented number, e.g. the nodes dropped in NodeDrop), thus it is OK to scale up on larger graphs. For Subgraph the complexity might be higher since there is a while loop in it, so when comes to too large graph, I think you might consider updating the implementation a little bit.

Because in the way of GraphCL, it needs 4 GNN encoders at the same time per epoch. Actually in GraphCL it only requires 1 GNN encoder, propagate for twice (2 augmented views). Things might be some different when comes to your case.

junkangwu commented 3 years ago

@yyou1996 Thanks a lot for your advice and suggestion, I think I have to check my code.