Causal Graph Synthesis

• Causal graph synthesis is the process of constructing and recovering causal models from observational, interventional, or mixed data sets.
• Methods range from combinatorial search and local partitioning to amortized neural approaches and visually grounded synthesis, each with unique guarantees.
• Synthesis targets vary from fully oriented DAGs and Markov equivalence classes to coarsened graphs and benchmark models, enabling diverse applications.

Causal graph synthesis denotes the construction, recovery, abstraction, or operational use of causal graph structure from observational data, interventional data, prior structural constraints, or learned data-to-graph mappings. In contemporary work, the synthesis target is not unique: it may be a directed acyclic graph (DAG), a Markov equivalence class represented by a CPDAG or PAG, an essential graph, a probability distribution over directed graphs, or a coarsened DAG whose nodes are groups of fine-grained variables. A closely related line of work uses a causal graph not as the output of learning, but as the fixed scaffold for synthetic data, image, or explanatory subgraph generation (Shah et al., 2024, Ke et al., 2022, Madaleno et al., 15 Jan 2026, Cheng et al., 17 May 2026).

1. Formal objects and synthesis targets

The common starting point is a DAG $G^*=(V,E^*)$ over variables $X=(X_1,\dots,X_p)$, with observational factorization

$$P(X_1,\dots,X_p)=\prod_{i=1}^p P\!\left(X_i \mid Pa^{G^*}(X_i)\right).$$

In observational settings, the synthesis target is often not a unique DAG but its Markov Equivalence Class (MEC), represented by a CPDAG $H^*$. A CPDAG encodes all DAGs with the same adjacencies and unshielded colliders as $G^*$. When latent variables are induced by subsetting or marginalization, MAGs and PAGs become the appropriate local or output objects rather than ordinary DAGs (Shah et al., 2024).

Different literatures synthesize different graph objects.

Object	Semantics	Representative use
DAG	Fully oriented causal graph	Score-based, neural, or benchmark generation
CPDAG / PAG / essential graph	Equivalence-class representation	Observational recovery, latent-variable settings
Directed adjacency matrix $A$	Full graph predicted entrywise	Amortized neural graph synthesis
Coarsened DAG	DAG over a partition of variables	Causal abstraction from interventional data

An essential graph is the standard equivalence-class object used in hybrid constraint-based/Bayesian search: it is a chain graph with directed and undirected edges, but no directed cycles, and it is sufficient to specify the set of undirected adjacencies and the set of $v$-structures (Dash et al., 2013). By contrast, amortized neural methods such as CSIvA target a full directed adjacency matrix $A$, with

$$A_{k,l}=1 \quad \text{iff there is an edge } X_l\to X_k,$$

so the predicted object is not an equivalence class but a fully directed graph, entry by entry (Ke et al., 2022). Recent abstraction work goes in a different direction again: a coarsening is a DAG $G'=(V',E')$ together with a surjection $X=(X_1,\dots,X_p)$0 such that

$X=(X_1,\dots,X_p)$1

making the output graph a causal model over grouped variables rather than the original feature set (Madaleno et al., 15 Jan 2026).

2. Search, decomposition, and direct graph construction

A central line of work treats causal graph synthesis as structured search under combinatorial and statistical constraints. In high-dimensional observational discovery, direct optimization over DAGs is NP-hard, so recent methods exploit a sparse undirected superstructure $X=(X_1,\dots,X_p)$2 satisfying

$X=(X_1,\dots,X_p)$3

and partition the variable set into overlapping subsets. In “causal graph partitioning,” a disjoint partition is expanded by its one-hop boundary, local causal discovery is run on each expanded subset, and a lightweight merge procedure, Screen, keeps an adjacency only if it appears in every local output containing that pair and then orients global unshielded colliders from local ones. Under the superstructure assumption, a consistent PAG learner $X=(X_1,\dots,X_p)$4, and a causal partition, the output recovers the true CPDAG in the infinite-data limit; the merge cost is

$X=(X_1,\dots,X_p)$5

and experiments report scalability to networks with up to $X=(X_1,\dots,X_p)$6 variables (Shah et al., 2024).

Other synthesis schemes reduce the search space differently. “From Causal Pairs to Causal Graphs” begins with pairwise probabilities $X=(X_1,\dots,X_p)$7 over $X=(X_1,\dots,X_p)$8, $X=(X_1,\dots,X_p)$9, or no edge, and defines a whole-graph model

$$P(X_1,\dots,X_p)=\prod_{i=1}^p P\!\left(X_i \mid Pa^{G^*}(X_i)\right).$$0

To enforce acyclicity, it builds a weighted matrix $$P(X_1,\dots,X_p)=\prod_{i=1}^p P\!\left(X_i \mid Pa^{G^*}(X_i)\right).$$1, approximates a maximum-likelihood topological order by topologically sorting the maximum spanning DAG of $$P(X_1,\dots,X_p)=\prod_{i=1}^p P\!\left(X_i \mid Pa^{G^*}(X_i)\right).$$2, and then restricts edges to point forward in that order, yielding the probabilistic and maximum-likelihood graph constructions PG, MLG, PDAG, and MLDAG. The resulting synthesis pipeline has polynomial runtime $$P(X_1,\dots,X_p)=\prod_{i=1}^p P\!\left(X_i \mid Pa^{G^*}(X_i)\right).$$3 in the number of variables (Rashid et al., 2022).

Hybrid search remains important when data are sparse or mixed-type. For mixed Gaussian and categorical variables, MGM-PCS and MGM-CPCS first estimate an undirected mixed graphical model and then run PC-stable or CPC-stable using a mixed-type likelihood-ratio conditional-independence test,

$$P(X_1,\dots,X_p)=\prod_{i=1}^p P\!\left(X_i \mid Pa^{G^*}(X_i)\right).$$4

thereby synthesizing a partially directed graph over mixed variables rather than forcing homogeneous modeling assumptions (Sedgewick et al., 2017). In sparse-data settings, EGS turns instability of PC into a search heuristic: it samples significance levels $$P(X_1,\dots,X_p)=\prod_{i=1}^p P\!\left(X_i \mid Pa^{G^*}(X_i)\right).$$5 and test orders, generates essential graphs, randomly converts each to a DAG, and scores candidates by

$$P(X_1,\dots,X_p)=\prod_{i=1}^p P\!\left(X_i \mid Pa^{G^*}(X_i)\right).$$6

yielding an anytime hybrid procedure over equivalence classes rather than a single deterministic PC output (Dash et al., 2013).

3. Learned, amortized, and visually grounded synthesis

A second major family replaces explicit graph search with learned data-to-graph maps. CSIvA, “Causal Structure Induction via Attention,” treats graph inference as supervised amortized posterior synthesis over DAGs conditioned on datasets. It learns

$$P(X_1,\dots,X_p)=\prod_{i=1}^p P\!\left(X_i \mid Pa^{G^*}(X_i)\right).$$7

an autoregressive Bernoulli distribution over the $$P(X_1,\dots,X_p)=\prod_{i=1}^p P\!\left(X_i \mid Pa^{G^*}(X_i)\right).$$8 entries of the adjacency matrix. The encoder operates on an $$P(X_1,\dots,X_p)=\prod_{i=1}^p P\!\left(X_i \mid Pa^{G^*}(X_i)\right).$$9 lattice, alternates attention across variables within a sample and across samples for a fixed node, and receives intervention metadata explicitly in an additional row. The decoder predicts the adjacency matrix row by row, entry by entry. The model outputs a full directed graph rather than a CPDAG or PAG, and acyclicity is not enforced during decoding (Ke et al., 2022).

In visually grounded settings, graph synthesis can be incremental rather than one-shot. ICIN, introduced for goal-directed control from RGB observations, synthesizes a weighted directed graph from intervention trajectories in which edges correspond to the likelihood that a switch controls a light. After encoding observations and computing state residuals, the graph is updated as

$H^*$0

so each observed transition contributes an attention-weighted edge update. The resulting graph is then used to contextualize a goal-conditioned policy, which attends to the relevant graph components when choosing actions (Nair et al., 2019).

These learned approaches shift the synthesis problem from combinatorial optimization to representation learning. They generally offer faster test-time inference and can exploit observational and interventional datasets jointly, but they also change the semantics of the output: the graph becomes the product of a learned conditional distribution over adjacency matrices or local causal structures, not a graph recovered by equivalence-class logic. This suggests a methodological split between “search over graph space” and “amortized graph posterior approximation,” with different guarantees and failure modes (Ke et al., 2022).

4. Abstraction, coarsening, and benchmark synthesis

Causal graph synthesis is not always aimed at the original variables. In causal abstraction, the objective is to synthesize a DAG over groups of variables. “Coarsening Causal DAG Models” defines a coarsening via a surjection $H^*$1, proves that the poset of valid coarsenings is a lattice and a sublattice of the partition refinement lattice, and introduces the interventional coarsening $H^*$2, where

$H^*$3

Under coarse Markov, coarse faithfulness, and interventional soundness, $H^*$4 is identifiable. The associated algorithm, RePaRe, directly learns the partition and abstract edge structure from multi-environment data with unknown intervention targets, and its worst-case runtime is

$H^*$5

for $H^*$6, $H^*$7, coarse graph size $H^*$8, maximal part size $H^*$9, and sample size $G^*$0 (Madaleno et al., 15 Jan 2026).

A distinct but closely related use of synthesis is benchmark generation. CRCG, the Causal Relationship Configurable Graph benchmark, synthesizes graph datasets with planted causal factors $G^*$1, confounders $G^*$2, and noise $G^*$3. It draws from a library of 25 motif types, connects motifs by four relation types—Adjacent, Cross, Entangled, and Containment—and induces train-test distribution shift by attaching spurious motifs conditionally on causal motifs during training and randomly at test time. The purpose is not causal discovery on the benchmark graph itself, but controlled generation of graph data in which causal, confounding, and noise components are known by construction (Gao et al., 2023).

Latent-confounded benchmark synthesis raises a further issue: many standard protocols encode hidden confounding only through off-diagonal entries of the idiosyncratic covariance matrix $G^*$4, often with diagonally dominant constructions that restrict the spectrum of correlation and partial-correlation matrices. “Addressing pitfalls in implicit unobserved confounding synthesis” replaces this with explicit latent-variable generation: $G^*$5 then hides selected variables and converts the resulting latent DAG into an ancestral ADMG for evaluation. The paper argues that block-hierarchical ancestral generation covers the implicit correlated-noise family while avoiding both diagonal-dominance spectral bias and unnecessary restrictions on valid confounded graph structures (Sun et al., 12 Mar 2025).

5. Graph-conditioned synthesis of data, images, and explanations

A neighboring meaning of causal graph synthesis uses a graph as a hard generative blueprint rather than as the primary learning target. CausalSynth begins from a Structural Causal Model

$G^*$6

samples low-dimensional causal skeletons by ancestral sampling in topological order, realizes them into high-dimensional outputs such as clinical notes or transaction logs with an LLM, and then applies iterative consistency verification. The skeleton distribution satisfies

$G^*$7

On ASIA, ALARM, and MIMIC-Struct, the method preserved conditional independencies with false-positive rates near the nominal $G^*$8 level and achieved realizability rates above 96% with 70B-parameter backbones. The contribution is therefore graph-faithful sample synthesis, not graph discovery (Cheng et al., 17 May 2026).

In medical imaging, the same logic appears in longitudinal causal image synthesis. A Tabular-Visual Causal Graph (TVCG) first fits structural mechanisms

$G^*$9

over demographics, biomarkers, and brain volumes, then couples the volume variables to a 3D image generator through an Intervened MRI Synthesis Module,

$A$0

This permits factual, interventional, and counterfactual MRI synthesis under irregular follow-up intervals, using the graph to propagate interventions on clinical variables into future anatomical images (Li et al., 2024).

At the explanation level, CXGNN constructs a local causal structure around a reference node, fits a graph-specific Neural Causal Model, estimates intervention effects of the form

$A$1

and returns the explanatory subgraph $A$2 induced by the node with maximal expressivity. Here the synthesized object is a causal explanatory subgraph rather than a full causal DAG, but the mechanism remains graph-conditioned and intervention-based (Behnam et al., 2024).

6. Domain-grounded construction, evaluation, and recurrent limits

Some work treats synthesis as graph authoring rather than automatic recovery. A human-centered augmented-reality framework for robotics supports variable selection, edge specification, graph-guided execution, interventions, and counterfactual-style refinement for a pick-and-place task, making causal graph creation a mixed-initiative engineering process instead of a purely statistical one (Tram et al., 2024). Text-derived pipelines define nodes as concepts and edges as extracted cause-effect relations, but assign each edge a probability distribution over a certainty factor $A$3, combining multiple textual qualifiers by grid-based multiplication and summarizing the posterior with KL-based adverb recovery; this yields uncertainty-weighted causal graphs rather than SCM-style DAGs (Garrido-Merchán et al., 2020).

Evaluation is correspondingly heterogeneous. Partition-based discovery reports TPR, SHD, FPR, and runtime (Shah et al., 2024). Pairwise aggregation uses TPR, FPR, and normalized SHD (Rashid et al., 2022). Amortized graph predictors such as CSIvA are usually evaluated by Hamming distance $A$4 between predicted and ground-truth adjacency matrices (Ke et al., 2022). Coarsened DAG learning uses ARI for partition recovery and F-score for edge recovery (Madaleno et al., 15 Jan 2026). Explanation systems emphasize exact groundtruth explanation identification (Behnam et al., 2024), while graph-conditioned generators evaluate conditional-independence preservation, realizability, KS $A$5-values, or downstream utility (Cheng et al., 17 May 2026, Li et al., 2024).

Several recurring constraints delimit what can be claimed. Divide-and-conquer MEC recovery depends on a valid superstructure $A$6, causal faithfulness, a consistent local PAG learner, and a causal partition (Shah et al., 2024). Interventional coarsening depends on coarse Markov, coarse faithfulness, and interventional soundness (Madaleno et al., 15 Jan 2026). Amortized methods inherit the inductive biases of their synthetic training distribution and may output a fully oriented adjacency matrix even when only an equivalence class is identifiable (Ke et al., 2022). Benchmark generators can narrow the supported model class if they rely on diagonally dominant covariance constructions or restricted bidirected-edge sampling (Sun et al., 12 Mar 2025). These contrasts imply that “causal graph synthesis” is not a single problem but a family of constructions whose outputs, guarantees, and failure modes differ sharply across MEC recovery, full DAG prediction, causal abstraction, benchmark generation, and graph-conditioned sample synthesis.

In a distinct theoretical usage, causal graph synthesis refers to the local realization of graph dynamics themselves. For vertex-preserving unitary causal dynamics over superpositions of bounded-degree graphs, every such global operator admits a finite-depth local decomposition

$A$7

where the $A$8 are commuting localized unitaries. In that setting, synthesis means constructive reduction of a global causal graph dynamics to local gates rather than recovery of a statistical causal structure (Arrighi et al., 2016).