First-order logic guides graph diffusion toward ontology-valid graphs
Original title: Logic-Guided Graph Denoising Diffusion Models
Fuzzy, differentiable rule relaxations steer a pre-trained denoiser at inference and raise constraint satisfaction.
The paper grounds graph diffusion models in domain knowledge expressed as first-order logic so generated graphs better meet structural constraints. Classifier guidance conditions sampling on a fuzzy, differentiable relaxation of the logical rules, applied at inference to a pre-trained model. Compared with unconditional knowledge-graph generation, adding domain ontologies as constraints significantly increases the share of graphs that satisfy them.
- Only FOL constraints need to be specified; the generative model stays pre-trained.
- Guidance optimizes rule satisfaction during the denoising trajectory.
- Unconditional diffusion can miss strict structural constraints and emit invalid graphs.
From the original article
1 min read at arxiv.orgGraph diffusion models are competitive approaches for generating complex graph-structured data. However, when such distributions have strict structural constraints, diffusion models may fail to capture them exactly, resulting in the generation of invalid graphs.
By Samuel Cognolato · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.