Research map / Label-efficient and robust learning / Test-time adaptation
Matcha: Mitigating Graph Structure Shifts with Test-Time Adaptation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Matcha, a plug-and-play graph test-time adaptation method that adapts GNN hop-aggregation parameters by minimizing a prediction-informed clustering loss, to handle graph structure shifts alongside existing TTA methods.
Problem
Generic test-time adaptation methods perform poorly on graphs under structure shifts (homophily and degree shifts), because structure shifts degrade node representation quality rather than only the classifier decision boundary.
Approach
Adapt only the hop-aggregation parameters (gamma) of a GPRGNN-style featurizer at test time by minimizing a label-free PIC loss (intra-class over total variance of soft pseudo-class clusters), which avoids the scale-driven trivial solutions of entropy minimization, and combine it with any base TTA method.
Key results
- Matcha alone improves source ERM accuracy by up to 31.95% on real-world datasets.
- When integrated with existing TTA methods, Matcha further boosts their performance by up to 22.72% on CSBM and up to 39.31% on real-world datasets.
- Adds only 11.9% extra computation per adaptation epoch versus 486% for GTrans and 247% for SOGA.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add Matcha to ERM by adapting hop-aggregation parameters | GPRGNN · CSBM homophily shifts, both directions | accuracy: 73.62 ± 0.44 -> 89.71 ± 0.27; 76.72 ± 0.89 -> 90.68 ± 0.26 | Helped |
| Add Matcha to ERM by adapting hop-aggregation parameters | GPRGNN · CSBM degree shifts, both directions | accuracy: 86.47 ± 0.38 -> 88.55 ± 0.44; 92.92 ± 0.43 -> 93.78 ± 0.74 | Helped |
| Add Matcha to ERM by adapting hop-aggregation parameters | GPRGNN · CSBM attribute + homophily shifts, both directions | accuracy: 61.06 ± 1.67 -> 85.34 ± 4.68; 72.61 ± 0.38 -> 74.70 ± 0.99 | Helped |
| Add Matcha to ERM by adapting hop-aggregation parameters | GPRGNN · CSBM attribute + degree shifts, both directions | accuracy: 77.63 ± 1.13 -> 78.29 ± 1.41; 73.60 ± 3.53 -> 73.86 ± 4.20 | Helped |
| Add Matcha to Tent | GPRGNN · CSBM, all eight shift settings | accuracy: 74.64 ± 0.38 -> 89.93 ± 0.16; 79.40 ± 0.57 -> 91.26 ± 0.08; 86.49 ± 0.50 -> 89.20 ± 0.20; 92.84 ± 0.18 -> 94.88 ± 0.09; | Helped |
| Add Matcha to T3A | GPRGNN · CSBM, all eight shift settings | accuracy | Helped |
| Add Matcha to AdaNPC | GPRGNN · CSBM, all eight shift settings | accuracy | Helped |
| Add Matcha to GTrans | GPRGNN · CSBM, all eight shift settings | accuracy | Helped |
| Add Matcha to SOGA | GPRGNN · CSBM, all eight shift settings | accuracy | Helped |
| Add Matcha to GraphPatcher | GPRGNN · CSBM, all eight shift settings | accuracy | Helped |
| Adapt only hop-aggregation parameters γ rather than MLP parameters θ (ablation) | Matcha on Syn-Products · Syn-Products | accuracy | Helped |
| Adapt only MLP parameters θ instead of hop-aggregation parameters (ablation) | Matcha on Syn-Products · Syn-Products | accuracy | Hurt |
Surprising findings
- Adapting both the MLP and hop-aggregation parameters leads to model forgetting, with accuracy rising then falling, whereas adapting only gamma is stable.
Limitations
- Convergence guarantee depends on the Lipschitz sensitivity of the base TTA algorithm.
- Gradients do not flow through the base TTA module, introducing gradient estimation error.
Experimental setup
Datasets: CSBM Syn-Cora Syn-Products Twitch-E OGB-Arxiv
Benchmarks: CSBM Syn-Cora Syn-Products Twitch-E OGB-Arxiv
Metrics: node classification accuracy
Baselines: ERM T3A Tent AdaNPC GTrans SOGA GraphPatcher
Builds on: GPRGNN Tent T3A AdaNPC SOGA
Closest papers by meaning
- Chasing Label Shifters: A Change-Aware Framework for Dynamic Graph Node Classification (NeurIPS 2026 · similarity 0.87)
- On the Benefits of Attribute-Driven Graph Domain Adaptation (ICLR 2025 · similarity 0.87)
- Pairwise Alignment Improves Graph Domain Adaptation (ICML 2024 · similarity 0.86)
- Smoothness Really Matters: A Simple Yet Effective Approach for Unsupervised Graph Domain Adaptation (AAAI 2025 · similarity 0.86)
- Test-time Adaptation on Graphs via Adaptive Subgraph-based Selection and Regularized Prototypes (ICML 2025 · similarity 0.85)
- Revisiting, Benchmarking and Understanding Unsupervised Graph Domain Adaptation (NeurIPS 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.