Graph is a Natural Regularization: Revisiting Vector Quantization for Graph Representation Learning
Researchers have developed RGVQ, a new framework to address codebook collapse in vector quantization for graph representation learning. This issue limits the expressiveness of graph data representations. RGVQ integrates graph topology and feature similarity as regularization signals, using soft assignments and structure-aware contrastive regularization to improve codebook utilization and token diversity. Experiments show RGVQ enhances performance across various downstream tasks, leading to more transferable graph token representations. AI
IMPACT Enhances graph representation learning, potentially improving performance in downstream AI tasks involving structured data.