Fine-Grained Graph Generation via Latent Mixture Scheduling

A new conditional variational autoencoder (CVAE) model for fine-grained structural control in graph generation has been introduced. Unlike existing methods that only offer coarse control over graph properties, this approach dynamically aligns graph- and property-driven representations in the decoder's latent space. A mixture scheduler progressively integrates graph and control priors to improve both graph fidelity and controllability. The model was tested on five real-world datasets and outperformed recent baselines in generation quality and control satisfaction. Applications include drug discovery, social network modeling, and knowledge graph construction.