Tags for DAGs: Graph Refinement with Meta-Informed Relations

Florian Peter Busch; Moritz Willig; Florian Guldan; Kristian Kersting; Devendra Singh Dhami

cs.LGJun 24, 2025

Tags for DAGs: Graph Refinement with Meta-Informed Relations

Authors: Florian Peter Busch, Moritz Willig, Florian Guldan, Kristian Kersting, Devendra Singh Dhami

Organizations: Computer Science Department, Technical University of Darmstadt · Hessian Center for AI (hessian.AI) · German Research Center for AI (DFKI) · Centre for Cognitive Science, Technical University of Darmstadt · Department of Mathematics and Computer Science, Eindhoven University of Technology

Abstract

Causal discovery has shifted from data-centric methods to hybrid strategies that integrate semantic knowledge from experts or large language models (LLMs). Such external information is vital for identifying causal structures beyond the Markov Equivalence Class (MEC), which data alone cannot resolve. However, expert availability is often limited, and LLMs frequently misidentify causal directions in specialized domains. To overcome such shortcomings, we propose a tag-based approach that leverages semantically meaningful labels while deriving causal directionality directly from data. Using variable-level tag assignments from available sources (e.g., LLMs), our tags for DAGs method learns from identifiable data structures to extract higher-level causal relations. These are then used to orient undirected edges, enabling causal discovery to move beyond the MEC without reliance on fallible external knowledge.

Figures & tables

Figure 1 : Collecting Tag Information. Our tagging approach considers the assignment of multiple tags per variable, (1) allowing for the flexible assignment of light-weight meta-information, and (2) repurposing already discovered directed edges for identifying higher-level causal relations.

Figure 2 : Tag-Supported Causal Discovery. The figure illustrates the tagging-based causal discovery described in this paper: First, a causal discovery algorithm such as PC or GES is used to obtain a CPDAG. Variables are tagged with semantically meaningful tags, for example, by using an LLM. Lastly, tag relations on the CPDAG are used to predict the direction of undirected edges.

	SHD	SHD double	SID min	SID max	Precision	Recall	F 1
	Ranks ( $\downarrow$ )	Ranks ( $\downarrow$ )	Ranks ( $\downarrow$ )	Ranks ( $\downarrow$ )	Ranks ( $\downarrow$ )	Ranks ( $\downarrow$ )	Ranks ( $\downarrow$ )
PC	$4.32{\scriptstyle\pm 0.38}$	$4.27{\scriptstyle\pm 0.33}$	$3.15{\scriptstyle\pm 0.28}$	$3.98{\scriptstyle\pm 0.48}$	$3.18{\scriptstyle\pm 0.39}$	$4.41{\scriptstyle\pm 0.34}$	$4.45{\scriptstyle\pm 0.31}$
GES	$3.54{\scriptstyle\pm 0.51}$	$3.07{\scriptstyle\pm 0.47}$	$1.27{\scriptstyle\pm 0.13}$	$3.17{\scriptstyle\pm 0.51}$	$2.69{\scriptstyle\pm 0.51}$	$3.17{\scriptstyle\pm 0.42}$	$3.13{\scriptstyle\pm 0.51}$
Typed-PC (Naive)	$3.42{\scriptstyle\pm 0.34}$	$3.62{\scriptstyle\pm 0.33}$	$3.10{\scriptstyle\pm 0.27}$	$3.20{\scriptstyle\pm 0.33}$	$3.07{\scriptstyle\pm 0.42}$	$3.47{\scriptstyle\pm 0.34}$	$3.59{\scriptstyle\pm 0.39}$
Typed-PC (Maj.)	$2.97{\scriptstyle\pm 0.28}$	$3.25{\scriptstyle\pm 0.34}$	$3.40{\scriptstyle\pm 0.38}$	$3.27{\scriptstyle\pm 0.37}$	$2.84{\scriptstyle\pm 0.35}$	$3.05{\scriptstyle\pm 0.26}$	$3.20{\scriptstyle\pm 0.24}$
Tagged-PC (AntiV)	$3.60{\scriptstyle\pm 0.45}$	$4.05{\scriptstyle\pm 0.38}$	$4.54{\scriptstyle\pm 0.32}$	$4.05{\scriptstyle\pm 0.46}$	$4.25{\scriptstyle\pm 0.40}$	$3.67{\scriptstyle\pm 0.37}$	$4.11{\scriptstyle\pm 0.40}$

Table 1 : Average Ranks for Best Configurations over all Datasets (lower is better). Tagging improves upon PC and GES, and both tagging methods outperform their typed counterparts. Tagged-GES achieves the best results overall, yielding the lowest ranks across all metrics. While the AntiV baseline offers a slight improvement over PC, it is surpassed by both typing and tagging approaches. Detailed absolute results are presented in Tab. in the appendix. ±y indicates standard deviation.

Figure 3 : Robustness to Faults. Tags generated by Claude 4.6. Smaller point and dashed lines indicate 0 edge directions (defaulted to 50%). Graph Faults (left): Removing edges has little effect on performance, whereas inverting edges has a stronger negative impact on accuracy. Tag Faults (right): Accuracy decreases with increasing noise, but remains substantially better than random chance. The dotted line shows an increasing abstention rate. Discussion on outliers (Cancer dataset) and other LLMs in Apps. and .

Figure 4 : Top: Tag Homogeneity for Claude 4.6. A tag relation has high homogeneity (blue/1) if variables of one tag consistently cause the other and low homogeneity (gray/0) if the cause-and-effect relationship between these variables is random. Blank fields have no evidence; low-evidence fields are hatched. Hatched areas might appear white on screen due to anti-aliasing. We observe high homogeneity throughout all datasets. See App. for high-resolution plots. Bottom: Tag Relations. Considering the most prominent tag pairs per dataset reveals high-level mechanisms present in the domain. We list tag pairs, with accuracy and support in brackets. Full list in App. .

Explore similar work

GENESIS: Towards Explainable Causal Discovery
Aug 4, 2026Abhinav Thorat, Ravi Kumar Kolla, Vishak K Bhat +2Causal Discovery Methods
User Engagement Research, Sony Research India
Causal Ensemble Agent: Hierarchical Causal Discovery with LLM-guided Expert Reweighting
Jun 9, 2026Xinyu Li, Yuanyuan Wang, Haoxuan Li +7Causal Discovery Methods Causal Graph
The University of Melbourne · MBZUAI · Peking University +4
DAG-FM: A Foundation Model for Causal Discovery under Heterogeneous Causal Mechanisms
Jul 13, 2026Yikang Chen, Zhengkang Guan, Haoyuan Qian +3Causal Discovery Methods Causal Modeling
1Zhejiang University · 2Tsinghua University

Aug 4, 2026, cs.LGJ/K move · Enter open · S save

GENESIS: Towards Explainable Causal Discovery

Abhinav Thorat, Ravi Kumar Kolla, Vishak K Bhat, Harsh Vardhan Singh Chauhan+1

User Engagement Research, Sony Research India

Causal Discovery Methods

Open paper AskPDF arXiv

Causal Discovery (CD) from observational data faces two fundamental challenges. First, purely statistical methods often lack the power to resolve structural ambiguities in low-sample regimes. Second, although LLM-assisted hybrid approaches improve structure recovery through semantic reasoning, the influence of that reasoning on individual edge decisions remains largely opaque. Consequently, existing hybrid methods fail to satisfy a fundamental requirement: explaining why a particular edge is included or excluded in the learned directed acyclic graph (DAG). This is critical in real-world applications, where no ground-truth DAG exists and every structural decision must be independently justified. We formalize this requirement as decision traceability, requiring every inferred edge to be supported by auditable statistical evidence, Markov Blanket consistency, or explicit domain reasoning. We propose GENESIS, an explainable hybrid CD framework that decomposes graph construction into interpretable decision points. GENESIS first identifies and scores three-node structural motifs, including chains, forks, and colliders, to establish transparent structural priors, then progressively refines the graph by integrating these priors with observational evidence, invoking domain knowledge only when statistical evidence is insufficient. By design, every edge decision is resolved through an auditable source of evidence. Experiments show that GENESIS achieves 100% decision traceability across all settings, establishing explainability as a first-class objective in causal discovery. Despite this additional requirement, GENESIS consistently outperforms purely statistical CD methods on the majority of benchmark datasets across all sample regimes in terms of Structural Hamming Distance (SHD), while achieving performance comparable to state-of-the-art LLM-assisted approaches.

Abstract

Explore similar work

GENESIS: Towards Explainable Causal Discovery

Causal Ensemble Agent: Hierarchical Causal Discovery with LLM-guided Expert Reweighting

DAG-FM: A Foundation Model for Causal Discovery under Heterogeneous Causal Mechanisms