1 of 19

Variational Inference for Training Graph Neural Networks in Low-Data Regime through Joint Structure-Label Estimation

Danning Lao *, Xinyu Yang *, Qitian Wu, Junchi Yan

Shanghai Jiao Tong University

2 of 19

What Is Graph?

Graph is an important representation of natural structures.

3 of 19

Graph Neural Network

4 of 19

Challenge: graph structure that can be used in aggregation is often incomplete

5 of 19

Challenge: node labels that can be used as�supervision can be very limited

6 of 19

GCN and GAT’s performance degrades notably with incomplete structure and limited labels

7 of 19

Can we design a weakly-supervised graph neural network jointly estimating graph structure and node label?

8 of 19

Formulate our problem setting

9 of 19

Variational inference for weakly-supervised graph learning

10 of 19

Two-branch architecture for label-structure joint inference

11 of 19

Results on Cora

12 of 19

Results on Citeseer

13 of 19

Results on Pubmed

14 of 19

Results under extreme low-data regime

15 of 19

Case study on Infectious Disease Spreading

16 of 19

Ablation study on model architecture

17 of 19

Node classification visualization

18 of 19

Link prediction visualization

19 of 19

Thanks for your attention!