Neural Causal Models

• Neural causal models are machine learning frameworks that blend structural causal models with neural networks to identify, estimate, and explain causal effects in complex, high-dimensional and nonparametric scenarios.
• They utilize diverse architectures—including SCM-parameterized networks and differentiable graph learning—to support observational, interventional, and counterfactual analysis with rigorous causal hierarchy principles.
• These models enhance AI interpretability and fairness by enabling counterfactual reasoning, causal attribution, and abstraction in data-intensive applications such as time series forecasting and neural explanation.

Neural causal models are a broad class of machine learning systems that integrate the structural principles of causal inference—most notably, structural causal models (SCMs) and directed acyclic graphs (DAGs)—with the expressiveness of neural networks. These models form a unifying framework for specifying, learning, and interrogating causal mechanisms in high-dimensional or nonparametric settings where classical approaches based on parametric structure or symbolic reasoning are limiting. Neural causal models support identification, estimation, and explanation of causal effects, both from observational and interventional data, and increasingly also enable abstraction, fair prediction, and counterfactual analysis in real-world domains.

1. Structural Definition and Foundations

At their core, neural causal models are parameterizations of structural causal models (SCMs) in which the structural equations (i.e., the mechanisms mapping parents and noise variables to each endogenous variable) are given by neural networks. Formally, let $V = \{X_1, ..., X_d, Y\}$ be observed random variables, and $G$ a DAG encoding the assumed (or learned) causal relationships. Each node $V_i$ is assigned a function: $$V_{i} = f_{i}(\mathrm{PA}_{V_{i}}, U_{i}), \quad U_{i} \sim P(U_{i})$$ where $\mathrm{PA}_{V_{i}}$ are the parents of $V_{i}$ in $G$, $U_{i}$ is exogenous noise, and each $f_{i}$ is a (potentially) neural network mapping. The joint distribution then factorizes as $P(V) = \prod_{i}P(V_{i} | \mathrm{PA}_{V_{i}})$ (Isozaki et al., 25 Dec 2025, Goudet et al., 2017, Xia et al., 2021).

This framework supports both classical forms—linear, polynomial—or fully nonlinear, nonparametric mechanisms via universal approximation, enabling the modeling of complex data-generating processes that depart from standard parametric assumptions.

2. Expressiveness, Learnability, and Causal Hierarchy

Neural causal models are universally expressive: any SCM with measurable noise can be approximated up to arbitrary accuracy by choosing sufficiently deep and wide neural architectures for the functions $f_i$ (Goudet et al., 2017, Xia et al., 2021). This supports not only matching observational ($L_1$), but also interventional ($L_2$) and counterfactual ($L_3$) distributions in the Pearlian causal hierarchy.

However, expressiveness does not guarantee learnability of causal effects or structures from data alone. The causal hierarchy theorem holds: no neural model, however complex, can identify general interventional or counterfactual queries from observational data unless the necessary structural constraints or interventional information are available. Imposing the correct inductive biases—via a user-supplied DAG, bidirected confounder cliques, or learned graph structure—is necessary for causal identification and estimation (Xia et al., 2021, Xia et al., 2022, Xia et al., 2024).

3. Model Classes, Training Objectives, and Algorithms

Neural causal modeling admits a range of architectural instantiations and algorithmic approaches:

• Direct SCM-parameterized NNs: Each mechanism $f_i$ is a feed-forward or recurrent neural network. Training proceeds by maximizing likelihood, adversarial objectives (e.g., Wasserstein GANs), or discrepancy minimization (e.g., MMD) between generated and real data (Goudet et al., 2017, Xia et al., 2022, Xia et al., 2024).
• Graph Structural Learning: Differentiable graph-parameterizations (e.g., via continuous adjacency matrices and acyclicity constraints using NOTEARS-style penalties), allow learning both parameters and graph structure, sometimes under interventions or active experimental design (Scherrer et al., 2021, Ke et al., 2019, Ke et al., 2020).
• Counterfactual Generation: Modeling counterfactuals involves abduction–action–prediction using neural abductor nets to infer exogenous states and regularization losses (e.g., MMD, kernel least squares) to enforce layer-$L_3$ (counterfactual) consistency (Kher et al., 18 Feb 2025).
• Abstraction and Representation Learning: Neural causal models now support variable and domain clustering as well as learned abstractions for high-dimensional and structured data (e.g., images), often coupled with representation-learning losses and cluster-wise encoders (Xia et al., 2024).
• Active Interventions and Fast Adaptation: Active intervention targeting uses the disagreement across sampled graphs and functional models to minimize the required number of interventions to identify the true DAG structure (Scherrer et al., 2021, Ke et al., 2020).

4. Causal Explanation, Attribution, and Interpretability

A central application domain for neural causal models is in causal attribution and explanation for neural network predictions. Frameworks such as CENNET augment neural predictors by extracting characteristic correlated variable sets (CCVs) through causal discovery algorithms, replacing purely correlational explanations by truly causal “reason sets” (Isozaki et al., 25 Dec 2025). Counterfactual causal attributions, e.g., average causal effect (ACE) and controlled direct effect (ACDE) of input features, are computed via intervention (do-operator) semantics applied to differentiable neural architectures (Chattopadhyay et al., 2019, Reddy, 2021).

Entropy-based explanation power indices (e.g., EMI, TEP) and pointwise mutual information quantify the causal explanatory strength of candidate variable sets, enabling ranking and selection of minimal predictive “reasons” that respect the underlying SCM (Isozaki et al., 25 Dec 2025).

Schemes for inducing causal structure in neural networks, such as Interchange Intervention Training (IIT), ensure that learned representations or modules correspond to aligned variables in symbolic SCMs, enhancing interpretability and systematic generalization (Geiger et al., 2021).

5. Applications: Identification, Estimation, Abstraction

• Causal Identification and Estimation: Given sufficient interventional data and a known or partially learned graph, neural causal models provide necessary and sufficient algorithms for identifying and estimating causal queries (ATE, ETT, NDE, etc.), with optimization-based bracketing characterizing identifiability (Xia et al., 2022, Xia et al., 2021).
• Causal Abstractions: Variable and domain clustering enables construction of higher-level causal models, allowing robust inference and identification at abstracted levels of granularity, with neural models learned directly on these abstractions (Xia et al., 2024).
• Counterfactual Fairness and Fair ML: Neural causal models provide mechanisms for generating counterfactually fair predictors by enforcing distributional alignment between factual and counterfactual outcomes of sensitive variables; explicit kernel least-squares losses are used to guarantee L₃-level fairness constraints (Kher et al., 18 Feb 2025).
• Time Series and Dynamical Systems: Neural Additive VAR (NAVAR) models uncover nonlinear Granger-causal structure in multivariate time series by parameterizing scalar contributions with neural nets, allowing ranking and interpretability of links in the resulting graphs (Bussmann et al., 2020). Extensions to causal market simulators and state-space fMRI models demonstrate the use of neural SCMs in generating counterfactual trajectories and large-scale causal brain networks (Thumm et al., 6 Nov 2025, Bae et al., 20 Oct 2025).

6. Limitations, Theoretical Guarantees, and Future Directions

Theoretical guarantees for neural causal models include universal approximation of SCMs (up to L₁–L₃ distributions) and identifiability results under suitable assumptions—such as injective mixers and sufficient distribution shifts for latent causal discovery (Liu et al., 2024). Nevertheless, identifiability can only be achieved when sufficient structural priors, interventions, or environmental shifts are available; otherwise, indistinguishable models at the observational level can encode radically different causal relations (Xia et al., 2021, Liu et al., 2024).

Limitations include:

• Scalability of search and estimation for large graphs or high-dimensional data (sample complexity, computational cost) (Goudet et al., 2017, Ke et al., 2020).
• Need for discretization or binning in certain explanation metrics (e.g., entropy in CENNET) (Isozaki et al., 25 Dec 2025).
• Challenges in counterfactual or fairness estimation due to abductor net approximation error, as L₃-consistency may not be perfectly enforced (Kher et al., 18 Feb 2025).
• Structure learning in the presence of hidden confounding, cycles, or unknown interventions remains an open challenge (Ke et al., 2019, Scherrer et al., 2021).
• For interpretable neural SCMs (e.g., TRAM-DAG), counterfactuals can only be directly computed in continuous settings due to invertibility constraints (Sick et al., 20 Mar 2025).

Open research areas involve extending neural causal models to integrate latent variables, handle nonparametric time series, learn abstractions jointly with SCMs in complex domains (images, graphs), and design scalable algorithms for structure learning under partial or uncertain interventions.

7. Empirical Results and Benchmarks

Empirical studies across diverse paradigms demonstrate the practical power of neural causal models:

• CGNNs attain >95% AUPR in multivariate causal discovery and cause–effect direction on real and synthetic benchmarks, outperforming classical competitors (Goudet et al., 2017).
• CENNET achieves top average rank of ground-truth causal variables and robustly eliminates pseudo-correlation in tabular data, often exceeding LIME/SHAP/ACV on both synthetic and real Bayesian networks (Isozaki et al., 25 Dec 2025).
• In fairness tasks, counterfactually fair NCMs with explicit L₃ enforcement realize better fairness–utility trade-offs versus prior adversarial-MMD methods (Kher et al., 18 Feb 2025).
• In time-series causal inference, NAVAR/LSTM outperforms PCMCI/GPDC/SLARAC and recovers nonlinearity and correct lag structure in both synthetic and gene expression networks (Bussmann et al., 2020).
• For market simulation, TNCM-VAE matches ground truth counterfactual probabilities with L₁ error ~0.06, outperforming standard VAE or GAN approaches (Thumm et al., 6 Nov 2025).
• In large-scale causal fMRI analysis, CausalMamba recovers correct neural connectivity with 37% higher causal accuracy and 88% pathway recovery rate versus DCM, scaling to hundreds of ROIs (Bae et al., 20 Oct 2025).
• For GNN model explanation, NCExplainer achieves up to 100% ground-truth match in synthetic datasets and 66–75% in real datasets, substantially exceeding associative GNN explainers (Behnam et al., 2024).

These results underscore that neural causal models, when appropriately structured, trained, and interpreted, provide a scalable, expressive, and theoretically principled foundation for causal inference, explanation, and robust prediction in modern data-intensive settings.