1 (Supplement slides for reading paper) Semi-Supervised Graph Classification: A Hierarchical Graph Perspective(WWW19) izunan385

Li, Jia, et al. "Semi-Supervised Graph Classification: A Hierarchical Graph Perspective." (2019).

• Task Collect Class Prediction for unlabeled

• input each graph instance: g labeled graph set and unlabeled graph set graph instance adjacency matrix

• output IC(graph Instance Classifier) receives graph info and outputs instance representation matrix predicted class probability vector HC(Hierarchical Graph Classifier) receives all graph instance( ) representation from IC graph-graph adjacency matrix and outputs predicted class prob matrix for all

• Task Collect Class Prediction for unlabeled • Loss function labeled graph instances unlabeled graph instances

• Supervised Loss (for labeled graphs ) • Disagreement Loss(for unlabeled graphs ) Disagreement means IC and HC prediction mismatch.

No content

GCN W0: learnable parameter

GCN with self loop W0: learnable parameter

GCN(summarized) 0 https://www.experoinc.com/post/node-classification-by-graph-con network Adjacent/co-occurrence matrix has structure information. Propagation rule is learned during training.

https://docs.dgl.ai/tutorials/models/1_gnn/9_gat.html

Cautious Iteration

Cautious Iteration Here, sampling top confident prediction for each step

Active Iteration Disagreement means IC and HC prediction mismatch. Ask annotator for annotating class of graphs which HC and IC have top-disagreement with.