Multi-Semantic Basis Enhances Graph Foundation Models for Multi-Label Tasks

Multi-label node classification, where nodes possess multiple simultaneous meanings, is a critical but challenging task in graph learning. While current methods can handle multiple labels, they often lack cross-domain generalization, meaning they perform poorly when applied to graphs outside their training domain. Graph Foundation Models (GFMs) offer a promising solution for transferable graph representations, but existing GFMs assume a single label per node, leading to "semantic entanglement" where multiple meanings are compressed into a single, less discriminative representation. To overcome these limitations, researchers propose the Multi-Semantic Basis Graph Foundation Model (MSB-GFM). This framework introduces a novel multi-semantic basis representation learning paradigm, allowing each multi-label node to be represented as an adaptive combination of distinct semantic bases. This provides a more flexible and accurate way to model multiple semantics. Furthermore, MSB-GFM incorporates a semantic-structure dual-channel architecture combined with domain adversarial training to facilitate effective knowledge transfer across different graph domains. Extensive experiments confirm the model's effectiveness.