Causal Graph Profiling

• Causal graph profiling is a method for characterizing graph-structured systems using explicit causal semantics rather than mere association.
• It integrates techniques such as structural causal modeling, uncertainty weighting, and estimand-based comparisons to enhance causal inference.
• Applications span anomaly detection, recommender systems, and text/database causal analysis, offering actionable insights into complex networks.

Searching arXiv for papers on causal graph profiling and closely related graph-causal methods. arXiv search query: "all: causal graph profiling" Causal graph profiling denotes a family of procedures that characterize graph-structured systems in explicitly causal terms rather than by association alone. Across the cited literature, the term ranges from fixing a causal graph and fitting a structural causal model with neural parameterizations (Wang et al., 2023), to weighting text-derived causal edges by probability distributions over certainty factors (Garrido-Merchán et al., 2020), to comparing graphs by the cause-effect estimands they support under latent confounding (Li et al., 28 Oct 2025), to learning regime-specific causal graph templates for anomaly detection (Malarkkan et al., 13 Aug 2025). It also includes environment-based profiling of which causal questions are identifiable in an observational study (Berzuini et al., 2024), and database-oriented representations in which causal variables, structural equations, and interventions are stored and queried as graph-native objects (Pachera et al., 2024).

1. Scope and principal senses of the term

Across the cited literature, causal graph profiling does not denote a single standardized algorithm. It refers instead to several related operations: specifying causal graph structure, quantifying uncertainty or invariance on that structure, learning graph-conditioned causal representations, and evaluating graphs by the causal queries they support. In all cases, the profiled object is not merely adjacency, but a graph endowed with causal semantics.

Usage	Profiled object	Representative papers
SCM-centered graph learning	structural equations, exogenous variables, interventional behavior	(Wang et al., 2023, Behnam et al., 2024)
Uncertainty-weighted causal graphs	edge certainty factors as probability distributions on $[0,1]$	(Garrido-Merchán et al., 2020)
Estimand-based graph comparison	sets of identifying expressions for $p(y \mid \operatorname{do}(t))$	(Li et al., 28 Oct 2025)
Regime or environment profiling	invariant causal structures across environments or operational states	(Malarkkan et al., 13 Aug 2025, Borriero et al., 11 Jun 2026, Berzuini et al., 2024)
Graph-system integration	hypernodes, structural equations, and causal query operators	(Pachera et al., 2024)

A recurrent distinction is between profiling a graph as a causal data structure and profiling a model through the graph. In recommendation and GNN explanation, the graph is embedded in a structural causal model whose parameters are learned (Wang et al., 2023). In text mining, the graph itself is the product, and profiling means characterizing the strength, uncertainty, and conflict of extracted cause-effect edges (Garrido-Merchán et al., 2020). In ADMG comparison, profiling means cataloguing which causal estimands a graph makes identifiable and how those estimands change under perturbations (Li et al., 28 Oct 2025).

2. Graph representations and causal semantics

One dominant formulation treats profiling as explicit structural causal modeling. In Neural Causal Graph Collaborative Filtering, the causal graph is defined as $G = (\{\mathcal{V}, Z\}, \mathcal{E})$, where $\mathcal{V} = \{U, V, E, Y\}$ are endogenous variables, $Z = \{Z_u, Z_v\}$ are exogenous variables, and the associated structural causal model is $$\mathcal{M} = \langle \mathcal{V}, Z, \mathcal{F}, P(Z) \rangle$$ with structural equations for $U$, $V$, $E$, and $Y$ (Wang et al., 2023). In that formulation, profiling means fixing the causal graph structure over users, items, preferences, outcomes, and exogenous factors, then learning embeddings that reflect those dependencies.

A second formulation uses mixed graphs rather than DAGs. “Graph Distance Based on Cause-Effect Estimands with Latents” works with acyclic directed mixed graphs (ADMGs), which contain directed edges $p(y \mid \operatorname{do}(t))$0, bidirected edges $p(y \mid \operatorname{do}(t))$1, and no directed cycles (Li et al., 28 Oct 2025). Directed edges encode possibly latent-mediated direct influences, while bidirected edges encode latent confounding between observed variables. Conditional ADMGs (CADMGs), districts, intrinsic sets, and fixing sequences then provide the semantics needed to characterize identifiability under hidden variables.

A third formulation puts probability distributions directly on edges. In “Uncertainty Weighted Causal Graphs,” the graph is $p(y \mid \operatorname{do}(t))$2, but each edge $p(y \mid \operatorname{do}(t))$3 carries a certainty factor $p(y \mid \operatorname{do}(t))$4, treated as a latent variable with a probability distribution rather than as a point estimate (Garrido-Merchán et al., 2020). Multiple textual observations of the same relation are combined by multiplying adverb-derived priors and normalizing on a grid,

$p(y \mid \operatorname{do}(t))$5

so profiling concerns the posterior shape of the edge certainty factor, not only the existence of the edge.

A fourth formulation is database-native. “What If: Causal Analysis with Graph Databases” defines a Causal Directed Acyclic Hypergraph as $p(y \mid \operatorname{do}(t))$6, where hypervertices represent causal variables, causal edges carry conditional and interventional probability distributions, and structural equations are attached to hypernodes (Pachera et al., 2024). This representation is designed so that causal variables, equations, interventions, and counterfactuals become first-class graph data objects.

3. Learning, inference, and computational mechanisms

A major computational pattern is to translate structural equations into trainable modules. In NCGCF, each structural function becomes a neural or probabilistic module, latent exogenous variables are Gaussian, and the neural causal model is learned with semi-implicit variational inference (Wang et al., 2023). The model includes causality-aware message passing,

$p(y \mid \operatorname{do}(t))$7

Gaussian encoders for users and items, and an ELBO that mixes clean and counterfactual preference scenarios through interventions on $p(y \mid \operatorname{do}(t))$8.

A related neural construction appears in GNN explanation. “Graph Neural Network Causal Explanation via Neural Causal Models” builds a causal structure $p(y \mid \operatorname{do}(t))$9 over node labels in a $G = (\{\mathcal{V}, Z\}, \mathcal{E})$0-hop neighborhood, proves the existence of a GNN-SCM and a $G = (\{\mathcal{V}, Z\}, \mathcal{E})$1-constrained GNN-NCM, and computes interventional probabilities $G = (\{\mathcal{V}, Z\}, \mathcal{E})$2 to rank explanatory subgraphs (Behnam et al., 2024). Profiling is then local and interventional: each node-centered neighborhood receives an expressivity score derived from the learned causal model.

Another mechanism is graph discovery under mixed or latent structure. “Mixed Graphical Models for Causal Analysis of Multi-modal Variables” supplies a mixed-type likelihood-ratio CI test and hybrid algorithms in which an undirected mixed graphical model restricts the search space for PC-stable or CPC-stable, yielding MGM-PCS and MGM-CPCS (Sedgewick et al., 2017). “Identification of Latent Variables From Graphical Model Residuals” instead diagnoses latent misspecification through residual structure: residual matrices are computed from a current DAG, PCA or an autoencoder estimates latent proxies $G = (\{\mathcal{V}, Z\}, \mathcal{E})$3, and the graph is relearned on $G = (\{\mathcal{V}, Z\}, \mathcal{E})$4 in an EM-like loop (Hayete et al., 2021).

Several works operationalize profiling through invariance. “Causal invariance in graphical models with latent variables” studies when $G = (\{\mathcal{V}, Z\}, \mathcal{E})$5 remains invariant or weakly invariant across environments after marginalizing latent parents, and for a multivariate Gaussian target proposes testing equality of residual covariance matrices across environments via Box’s $G = (\{\mathcal{V}, Z\}, \mathcal{E})$6 statistic (Borriero et al., 11 Jun 2026). “Causal Compression” uses directed information to select sparse subsets of time points that preserve causal flow between time series, with optimization of the form

$G = (\{\mathcal{V}, Z\}, \mathcal{E})$7

thereby profiling outgoing, incoming, and instantaneous causal flow in time (Wieczorek et al., 2016).

4. Profiling targets in major application domains

In recommender systems, the target is a user–item ecosystem whose graph structure is recast as exogenous factors, user and item variables, preference representation, and recommendation outcome. NCGCF profiles which neighbor relations are causally relevant for producing preferences, how exogenous factors such as conformity and exposure shape embeddings, and how interventions on the preference vector affect predicted interactions (Wang et al., 2023). In graph explanation, the target is a specific prediction: CXGNN profiles which nodes and edges are causally responsible for a graph label rather than merely associated with it (Behnam et al., 2024).

In graph representation learning for classification, profiling concentrates on invariant or causally meaningful substructure. CCAGNN disentangles causal and non-causal representations through a learned gate, mutual-information minimization, and intervention-based augmentation, so that node-feature dimensions with high gate values are treated as causal (Job et al., 20 Feb 2026). CNL-GNN performs structural-level interventions through Counterfactual Neighbourhood Generation, Edge Importance Estimation, adaptive edge perturbation, and context–object disentanglement, so that predictions rely on stable causal neighborhoods rather than spurious connections (Job et al., 20 Feb 2026).

In systems applications, the profiled object is often a regime or infrastructure state. CGAD follows a two-phase supervised framework—causal profiling and anomaly scoring—in which Dynamic Bayesian Networks are learned separately for “Normal” and “Attack” states, then compared by Structural Hamming Distance for each test segment (Malarkkan et al., 13 Aug 2025). CausIL profiles causal dependencies among microservice metrics by treating instances of a service as independent and identical conditioned on system assumptions, pooling instance-specific metric variations, and constraining score-based discovery with service-call and metric-type domain knowledge (Chakraborty et al., 2023). In the aneurysmal subarachnoid hemorrhage study, causal DAGs are used repeatedly during the study’s course to provide “midway insights,” determine which causal questions are identifiable and positive with current data, and guide targeted database enrichment (Berzuini et al., 2024).

In text and database settings, profiling is closer to causal knowledge management. Uncertainty-weighted text graphs profile the strength and imprecision of cause-effect statements such as “smoking $G = (\{\mathcal{V}, Z\}, \mathcal{E})$8 lung cancer” by maintaining a posterior distribution over each edge’s certainty factor (Garrido-Merchán et al., 2020). The graph-database vision paper profiles mediators, confounders, colliders, conditional and interventional probabilities, and counterfactuals through graph-native operators such as EXTRACT, PROBABILITY, and DO-CALCULUS (Pachera et al., 2024).

5. Evaluation, falsification, and benchmarking

A distinctive feature of this literature is that evaluation is often causal-task aligned rather than purely topological. In the ADMG setting, Fixing Identification Distance defines graph comparison through symbolic sets of valid estimands for $G = (\{\mathcal{V}, Z\}, \mathcal{E})$9. For graphs $\mathcal{V} = \{U, V, E, Y\}$0 and $\mathcal{V} = \{U, V, E, Y\}$1, with a verifier $\mathcal{V} = \{U, V, E, Y\}$2, the directional distance is

$\mathcal{V} = \{U, V, E, Y\}$3

and the normalized version averages this penalty over treatment–outcome pairs (Li et al., 28 Oct 2025). The point is not merely that edges differ, but that graph differences distort causal estimands, identifiability, or both.

A complementary perspective is falsification. “Toward Falsifying Causal Graphs Using a Permutation-Based Test” defines the set of local Markov condition violations $\mathcal{V} = \{U, V, E, Y\}$4, its fraction $\mathcal{V} = \{U, V, E, Y\}$5, and a permutation-based p-value

$\mathcal{V} = \{U, V, E, Y\}$6

which tests whether the graph is significantly better than random node permutations (Eulig et al., 2023). The same paper also introduces $\mathcal{V} = \{U, V, E, Y\}$7, the fraction of node permutations that are Markov equivalent to $\mathcal{V} = \{U, V, E, Y\}$8, as a structural measure of falsifiability.

Applied graph-learning papers use task-specific metrics. NCGCF reports Precision@K, Recall@K, and NDCG@K for $\mathcal{V} = \{U, V, E, Y\}$9, with ablations removing causality-aware message passing or counterfactual instance-aware ELBO (Wang et al., 2023). CXGNN uses graph explanation accuracy, graph explanation recall, and ground-truth match accuracy to evaluate whether a causal explanatory subgraph exactly matches a known motif or functional group (Behnam et al., 2024). CausIL evaluates adjacency precision, recall, F1, arrowhead precision, recall, F1, and Structural Hamming Distance on synthetic, semi-synthetic, and real microservice data (Chakraborty et al., 2023). CGAD uses point-adjusted F1-score, ROC-AUC, and PRC-AUC on segmented multivariate time series (Malarkkan et al., 13 Aug 2025). For mixed-variable causal discovery, precision, recall, MCC, and SHD are reported separately for cc, cd, and dd edge types (Sedgewick et al., 2017).

Benchmarking can itself be framed as profiling. CausalGraph2LLM evaluates LLMs with graph-level and node-level causal queries across seven graph encodings and finds that models are highly sensitive to the encoding used, with strong downstream effects on intervention tasks and additional biases when contextual information activates parametric memory (Sheth et al., 2024).

6. Assumptions, limitations, and directions

A recurrent limitation is that the causal graph is often assumed rather than discovered. NCGCF prespecifies directionality such as $Z = \{Z_u, Z_v\}$0, the role of $Z = \{Z_u, Z_v\}$1 and $Z = \{Z_u, Z_v\}$2, and the way exogenous variables enter; the paper explicitly notes that there is no causal discovery component, no proof of identifiability of specific effects, and interventions are performed on latent preferences rather than on observable actions or policies (Wang et al., 2023). Similar cautions apply to CCAGNN and CNL-GNN, whose causal or non-causal distinctions are learned at the representation level and depend on invariance-oriented inductive biases rather than on an explicitly validated SCM (Job et al., 20 Feb 2026, Job et al., 20 Feb 2026).

Latent-variable settings remain computationally and conceptually difficult. FID has exponential dependence on graph complexity through fixing-based identification, relies on symbolic canonicalization that can under-merge mathematically equivalent estimands, and currently targets interventional distributions $Z = \{Z_u, Z_v\}$3 rather than broader causal functionals (Li et al., 28 Oct 2025). Residual-based latent recovery improves structural inference of Gaussian graphical models and can be extended to ordinal and nonlinear cases, but theoretical guarantees are strongest in the Gaussian setting; for autoencoder variants the paper states that guarantees are not available, and predictive improvement is intrinsically capped relative to the confounded model (Hayete et al., 2021).

Other limitations concern uncertainty, scale, and data infrastructure. The uncertainty-weighted text-graph approach models edge certainty factors rather than structural equations or do-calculus, and explicitly distinguishes itself from Bayesian networks (Garrido-Merchán et al., 2020). The graph-database vision for causal analysis is deliberately conceptual: it specifies a causal extension of the property-graph model and candidate operators, but also identifies open problems in extraction, integration, maintenance, and efficient estimation inside a graph database (Pachera et al., 2024). In observational medicine, DAG-based profiling is productive precisely because it identifies when positivity is poor or key confounders are missing, but those same findings delimit what causal questions are presently answerable (Berzuini et al., 2024).

Taken together, these works suggest that causal graph profiling is best understood as a methodological layer above graph construction. It specifies what kind of causal object is being profiled—structural equations, edge-certainty distributions, invariant parent sets, regime-specific causal templates, estimand sets, or falsification statistics—and then asks how that object can be learned, queried, compared, or invalidated under the assumptions of the chosen framework.