Causal Subgraphs in Graph Representation Learning

• Causal subgraphs are minimal, invariant substructures that determine a graph's full predictive power by remaining unchanged under interventional shifts.
• Methodologies such as counterfactual-invariant GNN explainers and disentangled causal masking extract these subgraphs while blocking spurious correlations.
• Applications in graph classification, molecular property prediction, and reinforcement learning demonstrate improved accuracy and robust generalization.

Causal subgraphs are substructures within a graph that encode the irreducible set of edges, nodes, or higher-order motifs whose presence alone suffices to determine an outcome—typically a label or phenotype—under interventional and distributionally robust criteria. Unlike traditional GNN explanation methods that recover correlational or associative patterns, causal subgraph inference seeks minimal subgraphs that retain full predictive power under interventions on the remainder of the graph, block spurious correlations arising from confounding, and generalize across structural or distributional shifts. Causal subgraph inference has become central in graph machine learning, statistical causal inference, network biology, and knowledge-based causal reasoning.

1. Formal Definitions of Causal Subgraphs

The formalization of a causal subgraph varies with context, but key principles include sufficiency, minimality, and invariance under intervention. Given a graph $G = (V, E)$ and a label or phenotype $Y$, a causal subgraph $E^c \subseteq E$ is typically defined as the smallest subset of edges such that

$P(Y \mid E^c) = P(Y \mid E)$

and, under a stronger counterfactual-invariance condition,

$P(Y \mid \mathrm{CF}(E^s)) = P(Y \mid E),$

where $E^s = E \setminus E^c$, and $\mathrm{CF}(E^s)$ denotes resampling $E^s$ conditioned on $E^c$. This definition enforces that interventions (i.e., random or counterfactual replacement) on the non-causal subgraph do not change the distribution of $Y$ so long as $Y$0 is present (Zhang et al., 2024).

Alternative notions, particularly in causal graph structure learning, define a necessary and sufficient causal graph (NSCG) as the minimal subgraph whose G-factorization recovers the observed joint distribution and becomes insufficient if any variable is dropped (Cai et al., 2023). In out-of-distribution (OOD) generalization, the causal subgraph $Y$1 is the irreducible subgraph for which $Y$2 is invariant across environments $Y$3, and no strict subset shares this property (Qiu et al., 23 Oct 2025).

2. Methodological Frameworks for Causal Subgraph Discovery

Methodologies for inferring causal subgraphs are highly diverse and tailored to the problem context.

• Counterfactual-Invariant GNN Explainers: The CIDER framework builds a two-channel generative model, factorizing $Y$4, with dedicated variational encoders for causal and spurious subedges. A counterfactual-invariance loss ensures that causal subgraphs, once fixed, render interventions on the spurious subgraphs label-invariant. Diffusion-based distillation incrementally peels off non-causal edges (Zhang et al., 2024).
• Disentangled Causal Masking: Several works employ learnable (soft or hard) masks over edges or nodes, optimizing for shared, environment-invariant predictivity (causal) versus locally correlated, spurious predictions (spurious/confounding). Disentanglement is enforced by dual GNN branches with specialized causal and bias-aware loss functions, often enhanced with counterfactual mixing of representations (Fan et al., 2022, Wang et al., 21 Apr 2026).
• Information Bottleneck Objectives: Information-theoretic formulations extract minimal subgraphs maximizing $Y$5 while minimizing $Y$6 (Graph Information Bottleneck), with retrieval-based hybrid schemes combining motif clustering with learned compression (Rao et al., 2024, Li et al., 7 May 2025).
• IRM-Free Invariant Distribution Methods: By leveraging the principle that causal subgraphs are most distributionally stable across environments, norm-guided objectives iteratively optimize for subgraphs whose representations change least under cross-environmental distributional shift; representation norms serve as proxies for invariance (Qiu et al., 23 Oct 2025).
• Contrastive Learning for Confounding Control: In regression, enhanced information bottleneck criteria allow confounder subgraphs to retain some predictive power, while contrastive interventions (mixing causal with spurious representations from different samples) drive invariance and debiasing (Yin et al., 1 Jul 2025).
• Meta-Causal Graphs and Multi-Subgraph Worlds: In settings where the causal mechanism itself is meta-state-dependent, a finite set of causal subgraphs is indexed by latent world state, and curiosity-driven interventions drive exploratory learning of both meta-state partition and per-state subgraph (Zhao et al., 29 Jun 2025).
• Granger-Informed and Knowledge-Based Subgraphs: For temporal or knowledge-based settings, causal subgraphs are extracted via statistical tests of time-lagged influence (e.g., pairwise Granger causality in temporal social networks (Harit et al., 21 Jun 2025)) or via supervised subgraph mining and ranking within large-scale knowledge graphs for causal QA or language reasoning (Susanti et al., 10 Jun 2025, Zheng et al., 2022).

3. Theoretical Properties and Guarantees

Causal subgraphs admit rigorous theoretical characterizations:

• Invariance and Minimality: Under assumptions of invariant causal mechanism, the unique causal subgraph is the minimal set for which the conditional label distribution is environment-invariant and support is covered in unseen domains (Qiu et al., 23 Oct 2025).
• Identifiability under Markov Equivalence: For partial ancestral graphs (PAGs), subgraph-induction preserves ancestral and confounding structure, enabling recursive identification by bucket-decomposition and graphical invariance under the Markov equivalence class (Jaber et al., 2018).
• Bounds through Probabilities of Causation: Necessary and sufficient features are scored by probabilities of causation, providing lower bounds on natural total and direct effects (Cai et al., 2023).
• Relation to Norms and Representation Geometry: Cross-environment shifts in representation norms directly reflect departure from invariance, guiding subgraph selection (Qiu et al., 23 Oct 2025).
• Contrastive and Interventional Disentanglement: Random mixing or explicit permutation of spurious representations with causal codes can empirically and theoretically break d-connected bias paths, achieving effective deconfounding (Fan et al., 2022, Yin et al., 1 Jul 2025, Wang et al., 21 Apr 2026).
• Meta-State Disambiguation: When a system comprises multiple regimes (meta-causal subgraphs), identifying the latent partition is tied to recoverability of context-indexed causal structures (Zhao et al., 29 Jun 2025).

4. Algorithmic Implementations

Practical algorithms for causal subgraph extraction include:

Approach	Masking Paradigm	Intervention Strategy	Invariance Enforcement	Reference
CIDER (diffusion GNN)	VGAE dual-channel	Counterfactual resampling	Diffusion distillation	(Zhang et al., 2024)
Disentangled GNN	Edge/node soft masks	Mask/permute representations	Counterfactual+HSIC loss	(Fan et al., 2022)
RC-GNN (retrieval+GIB)	Hard subgraph	Cross-graph motif retrieval	Info-Bottleneck compression	(Rao et al., 2024)
IRM-Free Norm Guidance	Soft mask	Cross-environment maximization	Representation norm penalty	(Qiu et al., 23 Oct 2025)
Meta-Causal Agent	Causal codebooks	Active intervention selection	VQ, mask, curiosity loss	(Zhao et al., 29 Jun 2025)
CARNAS (causal NAS)	Edge scoring (MLP)	Embedding space mixing	Architecture consistency loss	(Li et al., 2024)

Implementation choices span variational inference (e.g., VGAE-style decoders, ELBO objectives), soft/hard masking via parameterized neural networks, discrete optimization through top-k edge selection, explicit statistical hypothesis tests (e.g., Granger), and meta-learning or reinforcement learning policies for dynamic interventions. Losses are composed to ensure both predictive sufficiency on $Y$7 and invariance of causal architecture under spurious interventions.

5. Applications and Empirical Results

Causal subgraphs have been empirically validated in numerous domains, achieving state-of-the-art performance in both accuracy and explainability under OOD and confounded settings.

• Graph Classification and Explanation: On motif-based benchmarks (e.g., BA-2motif), CIDER achieves ≈100% causal motif classification; on MUTAG and NCI1, causal subgraph identification provides 10-15 point accuracy improvements at high sparsity (Zhang et al., 2024). Retrieval-based GIB approaches boost precision/recall for explanatory subgraphs and enhance downstream prediction (Rao et al., 2024).
• Molecular Property and Drug Discovery: Causal subgraphs isolated by environment-aware objectives robustly generalize under scaffold and size splits (DrugOOD, OGBHIV, BBBP), outperforming both standard GNNs and rationale models (Li et al., 7 May 2025, Qiu et al., 23 Oct 2025).
• Node Classification in Heterophilic Graphs: Disentangling shortcut and causal subgraphs via counterfactual and independence regularization achieves top-1 or runner-up accuracy on seven highly heterophilic benchmarks (Wang et al., 21 Apr 2026).
• Causal Graph Regression: Soft-mask and contrastive-intervention models for regression tasks (GOOD-ZINC, ReactionOOD-SOOD) yield 20–40% MAE or 0.1–0.5 RMSE improvements over strongest baselines (Yin et al., 1 Jul 2025).
• Causal QA and Knowledge Discovery: Informative metapath subgraphs in biomedical KGs, ranked by supervised models and incorporated into LLM prompts, yield up to 44-point F1 improvement in knowledge-based causal discovery over non-subgraph or random baselines (Susanti et al., 10 Jun 2025).
• Meta-Causal Modelling: Identification of meta-state-driven causal subgraphs enhances model-based RL downstream reward and robustness to observation noise in both synthetic and robot manipulation environments (Zhao et al., 29 Jun 2025).

6. Limitations, Assumptions, and Future Directions

Current approaches to causal subgraph inference are subject to several critical assumptions and technical constraints:

• Model Assumptions: Many frameworks postulate additive noise models, two-channel structural causal models, or invariant mechanisms across environments; their validity must be empirically established (Zhang et al., 2024, Qiu et al., 23 Oct 2025).
• Hyperparameter Sensitivity: Edge sparsity constraints, mask/embedding dimension, batch size for contrastive interventions, and other architectural choices are not fully theoretically justified (Wang et al., 21 Apr 2026, Li et al., 2024).
• Identifiability: Theoretical uniqueness of the recovered causal subgraph (identifiability) can be challenging to establish, especially under limited data or overlapping motifs.
• Scalability and Efficiency: Enumeration of all subgraph candidates or interventions can be computationally expensive in large graphs or dense KGs, mitigated by sparsity-inducing penalties or efficient retrieval (Rao et al., 2024, Susanti et al., 10 Jun 2025).
• Extensions: Open research problems include handling higher-order or dynamic (temporal) subgraphs, extending to joint node- and graph-level causal subgraphs, incorporating adversarial confounder detection, and bridging to interpretability in deep neuro-symbolic systems (Zhang et al., 2024, Li et al., 7 May 2025).

A plausible implication is that integration of richer environment modeling, higher-order support (hypergraphs/modules), and explicit meta-state annotation will expand both the scope and the causal fidelity of causal subgraph frameworks in future work.

7. Connections to Broader Causal Inference and Graph Representation Learning

Causal subgraph inference bridges foundational causal graphical modeling, modern variational learning, and applied graph machine learning. Induced subgraphs in PAGs allow for identification under partial structural knowledge using sound graphical criteria (Jaber et al., 2018). Mask-based and intervention-based frameworks deeply connect to the literature on invariant representation, domain generalization, and information bottleneck theory. The necessary and sufficient causal graph paradigm provides a principled link to probabilities of causation, total/direct effects, and discrete feature selection under theoretical consistency guarantees (Cai et al., 2023).

Collectively, these developments define causal subgraphs as fundamental objects in the rigorous, interpretable, and robust analysis of relational and network data across scientific and data-driven domains.