Contrastive FUSE: Scalable Graph Learning with Partial Labels
Researchers introduced Contrastive FUSE, a framework for node representation learning in graphs with partial pairwise labels and no node features. It optimizes a spectral contrastive objective integrating community-aware signals with signed constraints. A lightweight approximation replaces the expensive modularity gradient, enabling efficient training on million-edge graphs. Experiments on benchmark citation networks, co-purchase graphs, and OGB datasets show competitive or superior classification performance.
Key facts
- Contrastive FUSE is a fast and unified framework for scalable node representation learning.
- It handles graphs with partially available pairwise node labels and no node features.
- The method directly optimizes a spectral contrastive objective.
- It integrates community-aware structural signals with signed pairwise constraints.
- A lightweight approximation replaces the expensive modularity gradient.
- The optimization scheme includes natural gradient decomposition and adaptive learning-rate scaling.
- Experiments were conducted on benchmark citation networks, large co-purchase graphs, and OGB datasets.
- Contrastive FUSE achieves competitive or superior contrastive classification performance.
Entities
Institutions
- arXiv