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

Efficient Memory Crystallization for Graph Learning under Non-Stationary Distribution Shifts

NeurIPS 2026 · Label-efficient and robust learning › Test-time adaptation · approach: Extrapolation and single-sample adaptation · 0 citations (OpenAlex)

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 EMC, a training-free test-time framework that crystallizes each incoming graph domain into a compact memory via a closed-form solution and models inter-domain dependencies with state-evolving memories, reducing runtime and GPU memory versus GCAL on continual graph shifts.

Problem

Graph learning under non-stationary (continually drifting) covariate shifts at test time, where source data and labels are unavailable, and prior replay/generative memory methods (e.g., GCAL) are computationally expensive and scale poorly.

Approach

Instead of training an auxiliary VGAE to generate memory graphs, crystallize each domain into a low-rank surrogate graph (~1% of nodes) in closed form by matching propagated features, and mix memories across consecutive domains via a convex state-evolving field, with a KL anti-forgetting term on stored memories.

Key results

• On Elliptic with GCN, EMC reduces runtime by 96.78% and GPU memory by 96.58% relative to GCAL (per paper's text).
• Averaged across datasets, EMC reduces runtime by 87.4% and GPU memory by 92.4% relative to GCAL.
• On Facebook-100 with GCN, EMC reaches AP 53.10 vs GCAL 52.72; on OGB-Arxiv with GCN, EMC AP 46.07 (SAGE backbone) vs GCAL 45.22 (SAGE).

Surprising findings

• STEM outperforms EMC on Twitch-Explicit AP (56.66 vs 55.95), which the paper does not dwell on.
• GCAL fails with OOM on GAT backbones for Twitch and Facebook, while EMC runs.

Limitations

• Memory size is a fixed fraction (1%) of node count, a hyperparameter choice not adaptively set.
• Theoretical bounds rely on assumptions (covariate shift, Lipschitz encoders, probability lower bound) that may not hold in practice.
• Reported baseline numbers for some methods are taken from published works rather than re-run.

Experimental setup

Datasets: Facebook-100 Twitch-Explicit OGB-Arxiv Elliptic

Benchmarks: Twitch-Explicit Facebook-100 Elliptic OGB-Arxiv

Metrics: AP AF TDG TDA runtime (s) GPU memory (MiB)

Baselines: DANN Norm TENT EERM GTrans LEBED CoTTA EATA GCAL STEM

Builds on: GCAL LEBED SGC Dirichlet energy regularization CoTTA

Closest papers by meaning

• GCAL: Adapting Graph Models to Evolving Domain Shifts (ICML 2025 · similarity 0.90)
• Towards Continuous Reuse of Graph Models via Holistic Memory Diversification (ICLR 2025 · similarity 0.88)
• GraphKeeper: Graph Domain-Incremental Learning via Knowledge Disentanglement and Preservation (NeurIPS 2025 · similarity 0.86)
• A Topology-aware Graph Coarsening Framework for Continual Graph Learning (NeurIPS 2024 · similarity 0.86)
• Towards Pre-trained Graph Condensation via Optimal Transport (NeurIPS 2025 · similarity 0.85)
• An Efficient Memory Module for Graph Few-Shot Class-Incremental Learning (NeurIPS 2024 · similarity 0.85)