Title: FLAGG: A Flexible Approach to Autoregressive Graph Generation

Abstract:

Current approaches to deep graph generation occupy two distinct ends of the spectrum: one-shot models, which produce nodes and edges simultaneously, and sequential models, which construct graphs through autoregressive sampling. While each methodology demonstrates superior performance in specific graph domains based on size and topology, neither is universally applicable. Specifically, one-shot techniques often falter when tasked with generating large-scale graphs, whereas sequential approaches tend to yield lower quality results on smaller graphs. To address these constraints, we introduce a system that flexibly integrates both methodologies.

We propose FLAGG (Flexible Autoregressive Graph Generation), a novel framework that employs one-shot models to sequentially generate segments of a graph. This architecture allows any one-shot model to function in an autoregressive manner, thereby providing the flexibility to select the sequential policy. This policy is defined via a stochastic node removal process, which an associated Insertion Model is trained to reverse. Our evaluation of FLAGG, utilizing the DiGress one-shot model across various datasets with differing graph sizes and domains, demonstrates that the proposed approach surpasses both standard one-shot and autoregressive baselines in terms of sampling quality.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC