Graph U-Nets Hongyang Gao and Shuiwang Ji Texas A&M University - - PowerPoint PPT Presentation

1 Graph U-Nets - Department of Computer Science & Engineering

Hongyang Gao and Shuiwang Ji

Graph U-Nets

Texas A&M University

2 Graph U-Nets - Department of Computer Science & Engineering

IMAGE VS. GRAPH

­ Image can be treated as a special graph with well-defined locality. There is no locality information on normal graph, which makes it hard to define pooling and un-pooling operation on graph data. ­ Node classification problems can be considered as image segmentation

• problems. Both predict for each node or pixel.

3 Graph U-Nets - Department of Computer Science & Engineering

U-NET ON GRAPH

Conv layer

­ GCN layer

Pooling layer

­ ?

Un-pooling layer

­ ?

https://lmb.informatik.uni-freiburg.de/people/ronneber/u-net/

Node classification Image segmentation

4 Graph U-Nets - Department of Computer Science & Engineering

GRAPH POOLING LAYER (GPOOL)

Projection Top k Node Selection Gate ! " # #

$ % %ℓ'( %ℓ

×

*ℓ'(

idx

*ℓ

Outputs top k sigmoid

⨀

Inputs 1 !

5 Graph U-Nets - Department of Computer Science & Engineering

GRAPH UN-POOLING LAYER (GUNPOOL)

gUnpool layer uses position information from gPool layer to reconstruct original graph structure.

gPool gUnpool GCN

6 Graph U-Nets - Department of Computer Science & Engineering

GRAPH U-NET

GCN

Inputs

GCN GCN GCN GCN gPool gPool gUnpool gUnpool Network Embedding

7 Graph U-Nets - Department of Computer Science & Engineering

NETWORK REPRESENTATION LEARNING RESULTS

Results on node classification tasks: Results on graph classification tasks:

8 Graph U-Nets - Department of Computer Science & Engineering

GRAPH U-NETS

Come to poster #25 for more details!