KDD 2023: Semi-Supervised Graph Imbalanced Regression 🔍

Gang Liu, University of Notre Dame

Molecules' graph structures reveal valuable insights into their properties. However, certain significant molecules often occupy a distinct corner, deviating from the majority of labels. Achieving equal prediction ability across all label areas is important. When we prioritize optimizing training errors in rare areas, there is a risk of compromising accuracy in popular areas. To overcome this challenge, we present SGIR: a semi-supervised framework to tackle the imbalanced regression problem on graphs. Our innovative approach sidesteps trade-offs by leveraging additional examples. Together, let's pave the way toward accurate and unbiased predictions.