Universal graph classifier supports varied features and prediction tasks
The Universal Classifier for Graph Learning
Machine Learning
Summary
Graph machine learning usually struggles because different graphs have different features and tasks. The authors introduce a single model called the Universal Classifier that can handle many types of graph data and predict nodes, edges, or whole graphs. It does this by turning diverse features and labels into a shared 3D space where similarity determines predictions. This lets the model work well even on new graphs it has never seen before and across different tasks without retraining. Their tests show it performs strongly on various graph learning challenges without needing task-specific adjustments.
What this means in practice
- •For data engineers: Use a single pretrained model to perform multiple graph prediction tasks across diverse datasets without retraining for each specific task.
- •For social network analysts: Predict user connections or community roles on networks with varying features using one unified classifier model.
Authors
Ben Finkelshtein, André Linhares, Petar Veličković, Bryan Perozzi, Mikhail Galkin
Abstract
While foundation models have revolutionized natural language processing and computer vision by leveraging universal vocabularies, Graph Machine Learning (GML) remains fractured due to the absence of a unified feature and structural representation across diverse domains. Existing works claiming to be Graph Foundation Models (GFMs) are typically restricted to node-level predictions or require fixed feature dimensions, failing to provide a truly task-agnostic backbone for the full spectrum of graph learning applications. In this paper, we introduce the Universal Classifier (UC), which supports arbitrary feature and class cardinalities, unifying node-, edge-, and graph-level objectives under a single similarity-based classification objective. The UC reformulates all node-, edge-, and graph-level prediction tasks as maximizing similarity in the latent space: by lifting heterogeneous features and labels into 3D latent tensors, the model learns transferable features independent of specific input schemas. This architecture allows a single pre-trained model to generalize to node classification, node regression, and link prediction across unseen graphs with varying feature semantics. Experiments show strong zero-shot transfer performance across node-, link-, and graph-level tasks.