Abstract: Graph generation poses challenges in handling large graphs due to their complexity. This paper presents a framework that effectively bridges existing graph compression methods with powerful generative models. Our key contribution lies in leveraging compression techniques and integrating them with the denoising diffusion model to address these challenges. By employing graph compression, we enable efficient diffusion processes and mitigate permutation problems. The framework facilitates the diffusion process by compressing the graph in terms of node count, allowing for the generation of large graphs with reduced computational complexity. Importantly, our approach ensures lossless decompression to preserve information during reconstruction. Experimental evaluation demonstrates the superior performance of our method compared to the base model, specifically on atom-level graphs. Our framework holds promise in advancing graph generation techniques and enabling their application across diverse domains, striking a balance between performance and efficiency while leveraging the power of the diffusion model.
0 Replies
Loading