To address the codebook collapse problem that arises when applying Vector Quantization (VQ) to learn discrete representations of graph-structured data, we propose the RGVQ framework, which integrates graph topology and feature similarity as explicit regularization signals. RGVQ enhances codebook utilization and token diversity through soft assignment via Gumbel-Softmax reparameterization and structure-aware contrastive normalization, thereby improving the performance of existing graph VQ-based models.
