Temel Kavramlar
Graph Beta Diffusion (GBD) is a novel generative model that leverages the flexibility of beta distributions to effectively capture the diverse statistical characteristics of graph data, including discrete structures and continuous node attributes, leading to improved realism in generated graphs.
İstatistikler
GBD achieved a degree MMD of 0.045 on the Grid benchmark, significantly surpassing all baselines.
On the Planar and SBM datasets, GBD achieved superior or comparable MMD scores on most graph statistics, along with high V.U.N. scores, consistently ranking first or second among all baselines.
GBD outperforms the basic continuous diffusion model (GDSS+TF) under the same GraphTransformer architecture on 2D molecule datasets.