I-CAM-UV: Integrating Causal Graphs over Non-Identical Variable Sets Using Causal Additive Models with Unobserved Variables

Hirofumi Suzuki, Kentaro Kanamori, Takuya Takagi, Thong Pham, Takashi Nicholas Maeda, Shohei Shimizu · 2026

Read I-CAM-UV: Integrating Causal Graphs over Non-Identical Variable Sets Using Causal Additive Models with Unobserved Variables by Hirofumi Suzuki, Kentaro Kanamori, Takuya Takagi, Thong Pham, Takashi Nicholas Maeda, Shohei Shimizu (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-03207.txt.

Introduction

Hirofumi Suzuki, Kentaro Kanamori, Takuya Takagi, and colleagues merge causal maps drawn from datasets that do not even share the same variables. Science often holds several studies of one system, each measuring a different subset of factors, with hidden confounders lurking. Learning one graph per dataset and overlapping them misses many real links.

Their method builds on causal additive models that retain information about unobserved variables in each graph. They prove the true causal graph stays structurally consistent with those per-dataset results. The integration algorithm then lists every graph compatible with all datasets, using an efficient combinatorial search. Tests against existing methods show the merged graphs recover relationships that overlap-based approaches cannot see.

Catalog record

CREATORHirofumi Suzuki, Kentaro Kanamori, Takuya Takagi, Thong Pham, Takashi Nicholas Maeda, Shohei Shimizu
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE50706
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.03207

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

I-CAM-UV: Integrating Causal Graphs over Non-Identical Variable Sets Using Causal Additive Models with Unobserved Variables Hirofumi Suzuki, Kentaro Kanamori, Takuya Takagi, Thong Pham, Takashi Nicholas Maeda, Shohei Shimizu Abstract Causal discovery from observational data is a fundamental tool in various fields of science. While existing approaches are typically designed for a single dataset, we often need to handle multiple datasets with non-identical variable sets in practice. One straightforward approach is to estimate a causal graph from each dataset and construct a single causal graph by overlapping. However, this approach identifies limited causal relationships because unobserved variables in each dataset can be confounders, and some variable pairs may be unobserved in any dataset. To address this issue, we leverage Causal Additive Models with Unobserved Variables (CAM-UV) that provide causal graphs having information related to unobserved variables. We show that the ground truth causal graph has structural consistency with the information of CAM-UV on each dataset. As a result, we propose an approach named I-CAM-UV to integrate CAM-UV results by enumerating all consistent causal graphs. We also provide an efficient combinatorial search algorithm and demonstrate the usefulness of I-CAM-UV against existing methods. Introduction Identifying causal relationships among variables is a fundamental task in various fields of science. While the most effective approach is a randomized controlled trial, such an approach is often difficult due to cost, ethical, and technical reasons. Thus, the mainstream is causal discovery from purely observational data ( 21 ) : estimating causal relationships as a causal graph, often assumed to be a directed acyclic graph (DAG). Recent works in causal discovery span diverse fields, including environmental science ( 16 ; 8 ) , biology ( 7 ; 19 ) , and materials and drug sciences ( 2 ; 29 ) . In many of the causal discovery examples, the number of variables is relatively small—often around 10—because of limits in what can be measured or easily understood. On the other hand, most of these studies assume that there are no unobserved confounders, even though such hidden common causes can exist even in small systems. This highlights the need for causal discovery methods that can handle unobserved confounding, especially in small to medium systems. Moreover, in many situations, multiple datasets share common objectives but differ in measurement settings such as observed variables. If we can leverage such datasets, it is expected to identify more causal relationships over just using a single dataset. Thus, in this study, we consider causal discovery from multiple datasets with non-identical variable sets ( 25 ; 13 ) . Despite the usefulness and demand, existing approaches are typically designed for a single dataset, assuming no unobserved confounders, and only a few approaches tackle the above scenario. ION ( 24 ) , IOD ( 26 ) , and COmbINE ( 27 ) construct partial ancestral graphs (PAGs) over the integrated variable set by utilizing conditional independence information. In contrast, CD-MiNi ( 10 ) constructs a full DAG by restricting the causal model to linear non-Gaussian cases. However, these approaches have the following limitations: PAGs contain uncertain information about causal relationships, and CD-MiNi is not suitable for non-linear cases. Aiming to estimate full DAGs and handle a more expressive model, we focus on the Causal Additive Model (CAM) ( 1 ) that captures non-linear causal relationships. One straightforward approach is to estimate a causal graph from each dataset by CAM and integrate them into a single graph. However, this approach identifies limited causal relationships because some variables in each dataset can be confounded with each other. Moreover, if some variable pairs are unobserved in any dataset, their causal relationships can not be identified by direct estimation from the datasets. To address this issue, we leverage the Causal Additive Model with Unobserved Variables (CAM-UV) ( 12 ) that allows for the presence of unobserved variables. CAM-UV provides structural information on causal graphs related to unobserved variables (unobserved causal path and unobserved backdoor path). We show that the ground truth causal graph has structural consistency with the provided information of CAM-UV on each dataset. Because multiple different DAGs satisfy such characterization, we propose an approach to enumerate possible DAGs and name the approach Integrating CAM-UV (I-CAM-UV). Moreover, we show an efficient algorithm for I-CAM-UV and confirm its usefulness via experiments. Our contributions are summarized as follows: • We propose an enumeration approach named I-CAM-UV to obtain integrated causal graphs of CAM-UV over multiple datasets with non-identical variable sets. • We show that I-CAM-UV recovers the ground truth DAG when CAM-UV results have no enormous errors and the integrated variable set has no further confounders. • We propose an efficient I-CAM-UV algorithm with a best-first strategy based on a monotonic inconsistency of DAGs and demonstrate its usefulness via experiments. Related Work While there are a variety of similar but slightly different datasets in the real world, almost all studies on causal discovery are specialized for a single dataset, and only a few studies tackle multiple datasets with non-identical variable sets ( 24 ; 27 ; 26 ; 10 ) . Our study challenges that field and aims to allow for richer and more practical causal analysis. Causal discovery methods can be categorized into three approaches: functional model-based ( 17 ; 9 ) , constraint-based ( 22 ; 23 ) , and score-based ( 3 ) methods. They are not mutually exclusive and can be combined. In this study, we focus on a non-linear functional model-based approach called the Causal Additive Model with Unobserved Variables (CAM-UV) ( 12 ) . Causal Additive Model (CAM) ( 1 ) assumes non-linear causal functions where effects from each variable can be separated linearly, which is a special case of Additive Noise Model (ANM) ( 9 ) . CAM-UV identifies causal relationships as much as possible and detects latent causal/backdoor paths containing unobserved variables. Leveraging the such ability of CAM-UV, we take a constraint-based approach to integrate CAM-UV results by a combinatorial search method. CD-MiNi ( 10 ) takes a functional model-based approach over multiple datasets with non-identical variable sets, similar to our method. It assumes a linear non-Gaussian acyclic model (LiNGAM) ( 17 ) and utilizes a continuous optimization algorithm based on the overcomplete independent component analysis to learn a causal graph ( 4 ) . While CD-MiNi handles linear causal relationships derived from LiNGAM, we handle non-linear causal relationships via CAM. ION ( 24 ) , IOD ( 26 ) , and COmbINE ( 27 ) focus on constraint-based approaches such as PC and FCI over multiple datasets with non-identical variable sets. They consider partial ancestral graphs (PAGs) representing Markov equivalence classes (MECs) of causal graphs and enumerate all possible PAGs having structural consistency with PC or FCI results. On the other hand, we consider enumerating not PAGs but full DAGs. That is, while the existing methods output a PAG set implicitly representing DAGs of MECs, our method outputs a single explicit DAG set. Preliminaries Let V=\{v_{i}\}_{i=1}^{d} be a variable set, and we consider causal relationships over V as a graph structure. Let G^{*}=(V,A^{*}) be the ground truth causal graph, which is a DAG with directed edges A^{*}\subset V^{2} . Each (v_{i},v_{j})\in A^{*} means v_{i} is a direct cause (parent) of v_{j} and

Read the complete text