Source: arXiv

Paper

Online Bayesian Node Classification on Inductive Graphs under Distribution Shift — arXiv 2609.13655 (Sep 2026).

What it claims

On evolving graphs, node classifiers must generalize inductively to newly arriving nodes under distribution shift and emit calibrated uncertainty for safety-sensitive calls. Standard GNNs are trained once and address neither — fine for frozen citation networks, brittle for live fraud or social graphs where vertices arrive daily.

The authors stack a Bayesian last-layer (BLL) on a deterministic GNN encoder. Categorical softmax breaks Gaussian conjugacy, so they introduce VBLL: a variational objective that jointly trains the encoder and an approximate last-layer posterior via an evidence lower bound with Monte Carlo expected log-likelihood. At test time they freeze the encoder and apply an online Laplace update to the last-layer posterior — a power-prior with exponential forgetting and a Kullback-Leibler anchor to the training posterior. The full method is online GVBLL.

Across five benchmarks under distribution shift, online GVBLL is the only method reported to rank first on both accuracy and negative log-likelihood on every dataset — up to 17 percentage points on Cora and 14 on ogbn-arxiv over the strongest non-GVBLL baseline, while staying competitive on calibration with MC Dropout, Deep Ensembles, Temperature Scaling, and Gaussian-process classifiers.

The breakdown

This is an online Bayesian head on a standard encoder, not a new GNN backbone — split variational training from Laplace streaming at deploy time. That matters when full-graph retraining is too slow.

The sweep is benchmark-bound: five shifted node-classification sets. 17 / 14 point jumps are dataset-specific existence proof, not a production fraud guarantee. Calibration is “competitive with” named baselines — open the PDF before wiring this into compliance slides.

Why readers outside the lab should care

Korean teams running transaction fraud, telecom abuse, or marketplace trust graphs see constant inductive arrivals while risk committees ask for confidence scores, not just labels. An online Bayesian head without full GNN retraining maps to that ops shape. Naver, Kakao, and cross-border recommendation graphs hit the same drift when viral content reshapes edges overnight.

Overseas readers on Korea-hosted graph stacks should not confuse Cora leaderboard gains with live pipelines — feature freshness and label delay still matter. If vendors sell “set-and-forget node classification,” streaming posterior updates on the last layer may be the cheaper lever before the next retrain cluster. Safety teams: ask whether production GNNs export NLL-ranked scores or only argmax labels.

What builders and Korea-touching teams should watch

• Do prototype online last-layer updates before committing to weekly full-GNN retrains on shifting social or payment graphs.
• Don’t treat 17 / 14 point benchmark lifts as guaranteed on your fraud or recommendation graph — verify on your inductive arrival rate and label skew.
• Expect encoder freeze at test time — feature pipelines must stay compatible; a silent encoder swap breaks the Laplace anchor story.
• Re-check calibration tables in the PDF against MC Dropout and Deep Ensembles on your class imbalance, not only citation benchmarks.
• Demand latency and memory numbers for the online Laplace step if you need sub-second scoring on high-degree nodes.

Context

Read this as a streaming Bayesian fix for GNNs on graphs that never stop growing — inductive nodes, distribution shift, and uncertainty in one package. Korelay’s frame: if your Korea-touching graph product only retrains offline, ask whether a GVBLL-style online head closes the accuracy and calibration gap before the next full retrain cycle.

Source

Primary: arXiv:2609.13655 (abstract and framing cited; open the OA PDF for methods and full tables). Do not republish the PDF.