Sparse Training of Discrete Diffusion Models for Graph Generation
This addresses scalability issues for researchers and practitioners working with large graph generation, though it appears incremental as it builds on existing diffusion models with a sparsity adaptation.
The paper tackles the quadratic complexity problem in generative graph models by introducing SparseDiff, a diffusion model that leverages sparse graph representations to achieve linear space complexity with selected edges. It demonstrates state-of-the-art performance across datasets and a fourfold speedup on the large Ego dataset compared to dense models.
Generative graph models struggle to scale due to the need to predict the existence or type of edges between all node pairs. To address the resulting quadratic complexity, existing scalable models often impose restrictive assumptions such as a cluster structure within graphs, thus limiting their applicability. To address this, we introduce SparseDiff, a novel diffusion model based on the observation that almost all large graphs are sparse. By selecting a subset of edges, SparseDiff effectively leverages sparse graph representations both during the noising process and within the denoising network, which ensures that space complexity scales linearly with the number of chosen edges. During inference, SparseDiff progressively fills the adjacency matrix with the selected subsets of edges, mirroring the training process. Our model demonstrates state-of-the-art performance across multiple metrics on both small and large datasets, confirming its effectiveness and robustness across varying graph sizes. It also ensures faster convergence, particularly on larger graphs, achieving a fourfold speedup on the large Ego dataset compared to dense models, thereby paving the way for broader applications.