DKCD: Domain Knowledge-Enhanced Causal Discovery from Unstructured Data

2026-07-13 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Health & Medical Research · Depth: Expert, extended

Summary

The DKCD (Domain Knowledge-enhanced Causal Discovery) framework addresses challenges in causal discovery from unstructured data within high-expertise domains like healthcare. Existing large language model (LLM) methods, such as COAT, are limited by insufficient identification of latent factors and unreliable factor annotations due to a lack of domain-specific knowledge and grounded reasoning. DKCD integrates domain knowledge graphs (KGs) with LLM-based reasoning through three interconnected components: Knowledge Mining, which retrieves relevant domain knowledge; Knowledge-guided Causal Reasoning, which discovers latent factors and generates causal clues for accurate annotation; and Causal Structure Discovery, which constructs final causal graphs. Experiments on two synthetic datasets (Diabetes and Respiratory), utilizing LLMs like GPT-4o, Gemini 2.5 Flash, Grok-3, and LLaMA 3-70B, demonstrate that DKCD significantly outperforms baselines in both causal factor identification and causal graph construction.

Key takeaway

For AI Scientists or Data Scientists building causal models from unstructured domain-specific text, you should integrate external domain knowledge graphs to overcome limitations of general LLMs. Your approach must explicitly address latent factor identification and reliable annotation, as demonstrated by DKCD's three-component framework. This will yield more complete and accurate causal graphs, but validate LLM-generated reasoning to mitigate hallucination risks.

Key insights

Integrating domain knowledge graphs with LLMs significantly enhances causal discovery from unstructured data in high-expertise fields.

Principles

• Domain knowledge is vital for latent factor identification.
• Knowledge-guided reasoning improves annotation accuracy.
• Complete factor sets yield more accurate causal graphs.

Method

DKCD extracts observable factors, retrieves relevant domain KGs, uses LLMs for knowledge-guided latent factor discovery and causal clue generation, then applies statistical algorithms like FCI for graph construction.

In practice

• Leverage domain KGs for implicit factor inference.
• Employ "causal clues" to guide LLM-based annotation.
• Utilize the FCI algorithm for robust causal graph inference.

Topics

• Causal Discovery
• Large Language Models
• Knowledge Graphs
• Unstructured Data Analysis
• Medical Informatics
• Causal Graph Construction

Best for: AI Scientist, Research Scientist, Data Scientist

Related on AIssential

• DKCD: Domain Knowledge-Enhanced Causal Discovery from Unstructured Data — DKCD improves causal discovery from unstructured data in high-expertise domains by integrating domain knowledge for latent factor identification and accurate annotation.
• Can an LLM Knowledge Graph Keep Two Unrelated Domains Apart? — The "myKG" pipeline effectively prevents LLM hallucination of cross-domain relationships in knowledge graph extraction through a multi-layered approach.
• Agentic Architecture: The Need For Knowledge Graphs & Structural Causal Models — The integration of Knowledge Graphs and Structural Causal Models is essential for building truly capable, trustworthy agents.
• Structure Learning on Clustered Data — A new method extends mixed models to DAG structure learning, enabling causal discovery in heterogeneous clustered data.
• CausalDS: Benchmarking Causal Reasoning in Data-Science Agents — CausalDS benchmarks LLM data-science agents on integrated causal reasoning, data analysis, and tool use, including abstention for non-identifiable queries.
• CausalForge: A Formally Grounded, Self-Improving Agentic Framework for Automated Research in Causal Inference — CausalForge automates theoretical causal inference research using a formally grounded, self-improving agentic framework to enhance reliability.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

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