Research map / Label-efficient and robust learning / Test-time adaptation

Test-time Adaptation on Graphs via Adaptive Subgraph-based Selection and Regularized Prototypes

ICML 2025 · Label-efficient and robust learning › Test-time adaptation · approach: Realistic and stable 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 ASSESS, a test-time adaptation method for graph classification that selects reliable test graphs via subgraph mutual-information-based adaptive thresholds and self-trains with regularized prototype supervision.

Problem

Test-time adaptation of graph neural networks under distribution shifts when training graphs are unavailable; existing TTA and GNN methods degrade on graph data and pseudo-labels are noisy.

Approach

Uses individual-level, structure-aware confidence thresholds derived from Jensen-Shannon mutual information between a graph and its subgraph, plus a Sinkhorn-balanced prototype self-training loss regularized toward prototypes initialized from the source classifier weights (a Bayesian MAP view).

Key results

• ASSESS achieves the best average accuracy on all five datasets, e.g. 78.9% on PROTEINS average versus 70.2% for RNA and 75.4% for GraphCL.
• On FRANKENSTEIN/Mutagenicity/NCI1 ablation, the full model averages 69.8% versus 68.0-68.7% when removing ASBS or either RPS term.
• ASBS selection accuracy reaches nearly 0.8 versus about 0.7 for class-wise thresholds on PROTEINS.

Measured effects reported in the paper

Surprising findings

• Unsupervised/semi-supervised methods like GraphCL and Mean Teacher are not consistently better than plain GNNs, and some TTA methods designed for Euclidean data underperform on graphs.

Limitations

• Theoretical analysis relies on strong assumptions (Tsybakov condition, PL condition, bounded gradients) and simplified selection approximations.
• Evaluated only on small TU graph classification benchmarks with split-by-density shifts.
• Relies on offline setting where all test inputs are available for adaptation.

Experimental setup

Datasets: FRANKENSTEIN Mutagenicity PROTEINS NCI1 IMDB-BINARY

Benchmarks: FRANKENSTEIN Mutagenicity PROTEINS NCI1 IMDB-BINARY

Metrics: classification accuracy selection accuracy

Baselines: GCN GraphSAGE GIN GAT Mean Teacher GraphCL SHOT TAST RNA

Builds on: SHOT Mean Teacher GraphCL TAST RNA Deep Graph Infomax-style Jensen-Shannon MI estimator

Closest papers by meaning

• Test-Time Graph Neural Dataset Search With Generative Projection (ICML 2025 · similarity 0.87)
• Revisiting, Benchmarking and Understanding Unsupervised Graph Domain Adaptation (NeurIPS 2024 · similarity 0.86)
• Matcha: Mitigating Graph Structure Shifts with Test-Time Adaptation (ICLR 2025 · similarity 0.85)
• Core Knowledge Learning Framework for Graph (AAAI 2025 · similarity 0.85)
• Towards Effective and General Graph Unlearning via Mutual Evolution (AAAI 2024 · similarity 0.85)
• Graph Data Selection for Domain Adaptation: A Model-Free Approach (NeurIPS 2025 · similarity 0.85)