# GABO: Graph Augmentations with Bi-level Optimization

**Authors:** Heejung W. Chung, Avoy Datta, Chris Waites  
**arXiv:** [2104.00722](https://arxiv.org/abs/2104.00722)  
**Date:** 2021-04-01  
**Summary page:** https://www.summarizepaper.com/en/paper/2104.00722v1/

## Key points

- Data augmentation techniques are used in machine learning to improve the generalization of models
- Automated methods have been developed for data augmentation that do not require domain knowledge about the dataset being used
- The study titled "GABO: Graph Augmentations with Bi-level Optimization" applied bilevel optimization to graph classification on the ogbg-molhiv dataset
- The best performing augmentation achieved a test ROCAUC score of 77.77% using a GIN+virtual classifier and outperformed state-of-the-art FLAG augmentation for this classifier on the leaderboard
- Bilevel optimization can be an effective tool for automating data augmentation and improving model performance in graph classification tasks without requiring extensive domain knowledge
- This study contributes to ongoing efforts towards developing more efficient and accurate machine learning models through automated data augmentation techniques.

## Summary

In the field of machine learning, data augmentation techniques are widely used to improve the generalization of models by augmenting training examples. Automated methods have been developed to address this problem as they do not require domain knowledge about the dataset being used. In this study titled "GABO: Graph Augmentations with Bi-level Optimization," authors Heejung W. Chung, Avoy Datta, and Chris Waites applied bilevel optimization - one such automated method - to graph classification on the ogbg-molhiv dataset. Their best performing augmentation achieved a test ROCAUC score of 77.77% using a GIN+virtual classifier and outperformed state-of-the-art FLAG augmentation for this classifier on the leaderboard. The framework combines a GIN layer augmentation generator with a bias transformation and demonstrates that bilevel optimization can be an effective tool for automating data augmentation and improving model performance in graph classification tasks without requiring extensive domain knowledge. This study contributes to ongoing efforts towards developing more efficient and accurate machine learning models through automated data augmentation techniques.