Generating Graph-Like Logical Rules for Knowledge Graph Reasoning via Diffusion Models

2026-06-06 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

GRiD is a novel framework that generates graph-like logical rules for Knowledge Graph (KG) reasoning, addressing limitations of traditional chain-like rule mining. It reformulates rule discovery as a discrete generative process using diffusion models, conditioned on the target relation. GRiD employs a two-phase training strategy: first, supervised pre-training captures structural priors from KG meta-graph subgraphs, then reinforcement learning fine-tunes the model using policy gradient optimization guided by non-differentiable rule-quality metrics like coverage and confidence. Experiments on six benchmark datasets, including YAGO3-10 and FB15K-237, demonstrate GRiD's competitive performance in KG completion. Ablation studies confirm its efficiency, robustness, and the complementary nature of graph-like rules, which are particularly beneficial on KGs with rich relational structures. The framework operates efficiently, with RL fine-tuning requiring 18.76 s per epoch on YAGO3-10 and peak GPU memory of 11.41 GB on FB15K-237.

Key takeaway

For AI Scientists and Machine Learning Engineers aiming to enhance Knowledge Graph reasoning, you should explore generative approaches like GRiD. This framework enables the discovery of complex, interpretable graph-like rules that overcome the limitations of traditional chain-like methods, especially in KGs with rich relational structures. Consider implementing GRiD's two-phase supervised pre-training and reinforcement learning fine-tuning to improve the accuracy and interpretability of your KG completion tasks.

Key insights

Graph-like logical rules for Knowledge Graph reasoning can be efficiently generated by combining discrete diffusion models with reinforcement learning.

Principles

• Graph-like rules offer richer relational patterns, complementing simpler chain-like rules in KGs.
• Supervised pre-training combined with reinforcement learning effectively optimizes non-differentiable rule quality metrics.
• Compact rule body sizes, such as S=6, provide sufficient expressiveness for complex graph-like rule patterns.

Method

GRiD reformulates rule discovery as a conditional discrete diffusion process over rule-body adjacency matrices, using SL pre-training for structural priors and RL fine-tuning with rule-quality metrics as rewards.

In practice

• When combining rule types, use a fusion weight of 0.1-0.2 for graph-like rules to maximize KGC performance.
• Employ the geometric mean of coverage, confidence, and PCA confidence as a robust reward signal for RL-based rule generation.

Topics

• Knowledge Graph Reasoning
• Diffusion Models
• Reinforcement Learning
• Logical Rules
• Graph Transformers
• Knowledge Graph Completion

Code references

• Haoxiang-Cheng/GRiD

Best for: Research Scientist, AI Scientist, Machine Learning Engineer

Related on AIssential

• GraphScout: Empowering Large Language Models with Intrinsic Exploration Ability for Agentic Graph Reasoning — GraphScout enables LLMs to autonomously explore knowledge graphs, internalizing reasoning through synthesized training data.
• Strikingness-Aware Evaluation for Temporal Knowledge Graph Reasoning — Evaluating Temporal Knowledge Graph Reasoning requires distinguishing rare, "striking" events from common ones.
• Knowledge Graph and Hypergraph Transformers with Repository-Attention and Journey-Based Role Transport — A novel architecture separates structured knowledge from language in transformers using a key-value repository and journey-based role transport.
• LLM Friendly Graphs for AI Memory — From Triples to Facts — Larger, context-rich labels and subgraph representations enhance LLM comprehension of complex facts in knowledge graphs.
• GS-Quant: Granular Semantic and Generative Structural Quantization for Knowledge Graph Completion — GS-Quant aligns graph embeddings with LLM tokens using semantically coherent, hierarchically structured discrete codes.
• BLINKG: A Benchmark for LLM-Integrated Knowledge Graph Generation — BLINKG benchmarks LLM capabilities in mapping diverse data to ontologies for Knowledge Graph generation.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.