Reproduce “Matcha: Mitigating Graph Structure Shifts with Test-Time Adaptation” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Wenxuan Bao, Zhichen Zeng, Zhining Liu, Hanghang Tong, Jingrui He

A framework called Matcha is introduced by the University of Illinois Urbana-Champaign to address performance degradation in Graph Neural Networks (GNNs) caused by graph structure shifts during test-time adaptation. The method, which adapts hop-aggregation parameters and uses a Prediction-Informed Clustering (PIC) loss, achieved performance improvements of up to 39.31% on real-world datasets and demonstrated high computational efficiency.

Key insights

  • Attribute shifts primarily impact the classifier's decision boundary, while graph structure shifts fundamentally degrade the quality of upstream node representations.
  • Adapting GNN's hop-aggregation parameters is crucial and effective for restoring distinguishable node representations when facing structure shifts.
  • A ratio-based Prediction-Informed Clustering (PIC) loss, focusing on intra-class compactness and inter-class separation, is more effective for improving representation quality than traditional entropy minimization.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.