Node4All: Single graph model beats 21 baselines across 25 datasets
No per-dataset tuning needed: a single model ranks 5th overall in node classification.
Node representation learning typically requires per-dataset training and hyperparameter tuning, limiting reusability. To solve this, Dooho Lee and Jaemin Yoo (accepted to KDD 2026) propose Node4All, a universal node representation learner built on two complementary ideas. First, the Channel Graph Transformer (CGT) uses a fixed parameterization that can process arbitrary graph datasets. Second, a self-supervised learning strategy trains on synthetic graphs, enabling generalization beyond any specific dataset. The model is evaluated on 25 node classification benchmarks against 21 supervised and self-supervised baselines. Remarkably, a single Node4All applied uniformly across all datasets ranks 5th among the 21 baselines—even though every baseline was individually optimized per dataset.
Beyond standard evaluation, Node4All supports one-shot and in-context learning with an appropriate predictor, outperforming recent Graph Foundation Models (GFMs) in these settings. This demonstrates that reusability across arbitrary graph datasets does not sacrifice effectiveness. The results open the door to graph models that can be deployed without expensive per-dataset tuning, similar to how large language models are used today. The authors have released code and model checkpoints, enabling others to build on this work.
- Node4All uses a Channel Graph Transformer (CGT) with fixed parameters that works on any graph dataset.
- Self-supervised learning on synthetic graphs enables generalization beyond individual datasets.
- Single Node4All ranks 5th out of 21 baselines on 25 benchmarks without any dataset-specific optimization.
Why It Matters
Enables reusable graph models without per-dataset tuning, accelerating graph ML deployment across industries.