[Paper Review] Integrating Large Language Model for Improved Causal Discovery

Taiyu Ban, L L Chen+5 more|arXiv (Cornell University)|Jun 29, 2023

Topic Modeling19 citations

TL;DR

The paper proposes a framework that combines knowledge-based LLM causal analysis with data-driven causal structure learning to recover more reliable causal structures from data, using LLM-informed ancestral constraints to guide CSL algorithms. It benchmarks multiple LLMs and CSL methods across diverse datasets, showing significant improvements in several cases while highlighting challenges and error tolerance needs.

ABSTRACT

Recovering the structure of causal graphical models from observational data is an essential yet challenging task for causal discovery in scientific scenarios. Domain-specific causal discovery usually relies on expert validation or prior analysis to improve the reliability of recovered causality, which is yet limited by the scarcity of expert resources. Recently, Large Language Models (LLM) have been used for causal analysis across various domain-specific scenarios, suggesting its potential as autonomous expert roles in guiding data-based structure learning. However, integrating LLMs into causal discovery faces challenges due to inaccuracies in LLM-based reasoning on revealing the actual causal structure. To address this challenge, we propose an error-tolerant LLM-driven causal discovery framework. The error-tolerant mechanism is designed three-fold with sufficient consideration on potential inaccuracies. In the LLM-based reasoning process, an accuracy-oriented prompting strategy restricts causal analysis to a reliable range. Next, a knowledge-to-structure transition aligns LLM-derived causal statements with structural causal interactions. In the structure learning process, the goodness-of-fit to data and adherence to LLM-derived priors are balanced to further address prior inaccuracies. Evaluation of eight real-world causal structures demonstrates the efficacy of our LLM-driven approach in improving data-based causal discovery, along with its robustness to inaccurate LLM-derived priors. Codes are available at https://github.com/tyMadara/LLM-CD.

Motivation & Objective

• Motivate integrating prior causal knowledge from LLMs with statistical CSL to address NP-hardness and data limitations.
• Quantify how LLM-derived causal statements can act as ancestral constraints to improve causal structure learning.
• Compare hard and soft constraint approaches for incorporating LLM priors into CSL across multiple datasets and LLMs.
• Assess which LLMs provide high-quality priors and where their guidance most improves CSL performance.

Proposed method

• Define a universal prompt set to extract direct causal relations from LLMs about given variables.
• Use a three-stage LLM process (Understanding, Causal Discovery, Revision) to produce a set S of direct causal statements.
• Convert S into ancestral constraints A for CSL, and integrate them via hard or soft constraint methods.
• Formalize hard constraints as pruning and order-based search to satisfy ancestral edges.
• Formalize soft constraints by augmenting the CSL score with a prior term reflecting the LLM guidance (Bayesian prior).
• Evaluate LLM+CSL against purely data-driven CSL across eight BN-repository datasets with multiple CSL algorithms and LLMs.

Experimental results

Research questions

• RQ1 Can LLM-driven causal statements improve the quality of recovered causal structures when combined with data-driven CSL?
• RQ2 How does the quality of LLM priors affect CSL performance across datasets and algorithms?
• RQ3 When is a hard constraint approach preferable to a soft constraint approach for incorporating LLM priors?
• RQ4 Which LLMs provide reliable causal priors for CSL, and where do they fail?

Key findings

• LLM priors can substantially improve CSL in many datasets, achieving noticeable F1 gains and SHD reductions compared to data-only methods.
• GPT-4 tends to produce high-quality qualitative causal statements on some datasets (e.g., Cancer, Asia) and improves integrating frameworks in those cases.
• Soft constraint methods generally show tolerance to some erroneous priors and can yield improvements even when priors are imperfect.
• Hard constraints excel when LLM priors are highly accurate, by aggressively pruning inconsistent BN structures.
• GPT-4’s qualitative accuracy varies by dataset and scale, with better performance on smaller/less parameter-heavy structures and weaker results on more complex ones.