Generalized Graph Variational Autoencoders: Bounded Divergences Control Posterior Collapse

Generalized Graph Variational Autoencoders (GGVA) replace the KL divergence in the VGAE objective with any R\'enyi-Tsallis divergence of order q. By selecting q < 1, the Tsallis divergence remains bounded by 1/(1-q) regardless of latent width, preventing posterior collapse in graph generative tasks.