GRiD: Diffusion Models for Graph-like Rule Mining in Knowledge Graphs
A new framework called GRiD reformulates graph-like rule mining for knowledge graph reasoning using diffusion models. Existing methods focus on simple chain-like rules, missing richer structures like cycles and branches, and face computational bottlenecks from combinatorial explosion. GRiD addresses these by aligning diffusion model training with rule quality metrics, enabling the generation of complex graph-like rules. The approach is detailed in arXiv:2605.30747.
Key facts
- GRiD uses diffusion models for graph-like rule mining.
- Existing rule mining methods focus on chain-like rules.
- Graph-like rules include cycles and branches.
- Combinatorial explosion challenges graph-like rule search.
- Diffusion models' training objectives misalign with rule quality.
- Non-differentiable KG rule metrics hinder optimization.
- GRiD reformulates rule mining to align with diffusion models.
- The paper is on arXiv with ID 2605.30747.
Entities
Institutions
- arXiv