# Graph U-Nets

**Authors:** Hongyang Gao, Shuiwang Ji  
**arXiv:** [1905.05178](https://arxiv.org/abs/1905.05178)  
**Date:** 2019-05-11  
**Summary page:** https://www.summarizepaper.com/en/paper/1905.05178v1/

## Key points

- "Graph U-Nets" paper addresses representation learning for graph data
- Challenges of using CNNs for graph data
- Introduction of novel operations: gPool and gUnpool
- gPool layer selects nodes based on scalar projection values to form a smaller graph
- gUnpool layer restores original structure using position information from selected nodes in gPool layer
- Development of Graph U-Nets model combining CNNs with adaptations for graph data
- Improved performance in node classification and graph classification tasks compared to previous models
- Incorporation of ideas from image pixel-wise prediction tasks
- Effective pooling and up-sampling operations on graphs

## Summary

, , , , 

The paper "Graph U-Nets" by Hongyang Gao and Shuiwang Ji addresses the problem of representation learning for graph data. While convolutional neural networks (CNNs) are effective for operating on images, they face challenges when dealing with graph data. The authors propose a solution to this problem by introducing novel operations called graph pooling (gPool) and unpooling (gUnpool). These operations aim to overcome the limitations of traditional pooling and up-sampling operations that are not well-suited for graphs. The gPool layer selectively chooses nodes based on their scalar projection values on a trainable projection vector to form a smaller graph. On the other hand, the gUnpool layer restores the original structure of the graph using position information from nodes selected in the corresponding gPool layer. Building upon these proposed layers, the authors develop an encoder-decoder model known as Graph U-Nets, which combines CNNs with adaptations for effective operation on graph data. To evaluate their approach, experimental results on node classification and graph classification tasks consistently demonstrate improved performance compared to previous models. This paper introduces an innovative solution for representation learning on graph data by incorporating ideas from image pixel-wise prediction tasks. The proposed gPool and gUnpool operations enable effective pooling and up-sampling operations on graphs, leading to improved performance in various classification tasks.