Towards Multi-Label Graph Foundation Models: From Single-Vector Representation Learning to Multi-Semantic Basis Learning

## Abstract Multi-label node classification is a critical yet challenging task in graph learning, where each node may simultaneously exhibit multiple semantic labels. Existing methods for this task often model multiple labels effectively but are confined to in-domain scenarios, requiring training and testing within the same graph domain, which limits their cross-domain generalization. Recently, Graph Foundation Models (GFMs) have emerged as a promising paradigm for learning transferable graph representations across diverse graph domains and downstream tasks. However, current GFMs are built on a single-label assumption, where each node is arbitrarily treated as containing only one semantic class and is embedded into a single vector representation. For multi-label nodes, such a representation approximates multiple semantics with a single point in the representation space, inevitably leading to semantic entanglement and making simultaneous discrimination of multiple labels difficult. To address these limitations, we propose the Multi-Semantic Basis Graph Foundation Model (MSB-GFM), a framework designed for cross-domain multi-label node classification. Specifically, we introduce a multi-semantic basis representation learning paradigm that models each multi-label node as an adaptive composition of semantic bases, thereby enabling flexible representational capacity to capture multiple semantics. Furthermore, we develop a semantic-structure dual-channel architecture with domain adversarial training to facilitate effective cross-domain knowledge transfer. Extensive experiments demonstrate the effectiveness of our proposed model.

via ArXiv AI