Skip to content

CausalGraphX: A Counterfactual Graph Neural Network Framework for Explainable Systemic Risk Assessment

🕒 Published (v1): 2026-07-15 23:10 UTC · Source: Arxiv · Venue: AAAI 26 · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

CausalGraphX combines Graph Attention Networks with adversarial causal regularization and counterfactual explanation generation to predict cascading bank defaults while identifying minimal, actionable interventions to prevent systemic crises. It addresses the dual failure of standard GNNs: black-box opacity and spurious-correlation learning. Validated on large-scale synthetic financial networks, it outperforms all baselines in predictive accuracy and produces sparse, causally-grounded counterfactual explanations.

Problem

Standard GNNs applied to systemic financial risk prediction (a) learn spurious correlations rather than causal shock-propagation pathways, making them fragile under interventions, and (b) provide no interpretable rationale for predicted defaults, which is unacceptable for regulatory stress-testing that requires "what-if" causal reasoning.

Method

CausalGraphX has two stages:

1. Causal Representation Learner (GAT + adversarial regularization): A 3-layer Graph Attention Network computes attention coefficients \(\alpha_{ij}\) incorporating edge weights \(W_{ij}\) (bilateral exposures) to produce node embeddings \(Z\). A minimax adversarial objective is added: $\(\min_\theta \max_\phi \; \mathcal{L}_\text{pred}(\theta) - \lambda \mathcal{L}_\text{adv}(\theta, \phi)\)$ where a discriminator \(D_\phi\) tries to predict a confounder \(C\) (e.g., regional identifier) from \(Z\), and the GNN is trained to make \(Z\) uninformative about \(C\), forcing reliance on causal pathways.

2. Counterfactual Explanation Generator: Given a predicted default, an Adam-based optimization finds a minimal perturbation \((X', W')\) to the feature matrix and adjacency weights that flips the prediction to survival: $\(G' = \arg\min_{X', W'} \mathcal{L}_\text{valid} + \gamma_1 \mathcal{L}_\text{dist} + \gamma_2 \mathcal{L}_\text{sparsity}\)$ with \(\mathcal{L}_\text{dist}\) using MAD-weighted \(\ell_2\) and \(\mathcal{L}_\text{sparsity}\) using \(\ell_1\) to favor interpretable, sparse interventions. Gradients backpropagate through the frozen GNN into the input space.

Networks are synthetic (Erdős–Rényi and Barabási–Albert topologies, \(N \in \{500, 1000, 2000\}\)), with ground-truth default cascades generated via a modified Eisenberg–Noe clearing simulator.

Key Contributions

  • CausalGraphX framework unifying GAT-based systemic risk prediction with counterfactual explainability.
  • Adversarial causal regularization (minimax, \(\lambda\)-weighted) that enforces confounder-invariant embeddings in financial graph settings.
  • Gradient-based counterfactual optimization over both node features and edge weights simultaneously (vs. prior work on subgraph selection only).
  • Empirical validation on synthetic financial networks with explicit confounders and known ground-truth causal mechanisms, enabling causal plausibility measurement.

Results

Predictive performance (Table 1, AUC / F1, ER network N=2000): - CausalGraphX: 0.921 / 0.871 - GAT (standard): 0.889 / 0.834 - GCN: 0.856 / 0.798 - DebtRank: 0.778 / 0.715 - Logistic Regression: 0.738 / 0.671

BA network N=2000 (Table 2): - PR-AUC: CausalGraphX 0.516 vs. GAT 0.428; MCC: 0.701 vs. 0.637

Counterfactual quality (Table 3, BA N=2000) vs. GAT: - Validity: 99.1% vs. 97.3% - Proximity (L2): 0.139 vs. 0.168 (−17.3%) - Sparsity (L0 features changed): 2.94 vs. 4.82 (−39.0%) - Causal plausibility: 89.3% vs. 61.7% (+44.8%) - Actionability: 91.6% vs. 72.4% (+26.5%)

Robustness (Table 5): AUC drop under adversarial attack: −0.072 (CausalGraphX) vs. −0.124 (GAT).

Case study: A $4.2B capital injection to Bank B (identified by counterfactual) reduces systemic losses from $142B to $31B — a 33.8× return on intervention cost.

Limitations

  • Evaluated exclusively on synthetic/simulator-generated data; no validation on real-world interbank exposure datasets (e.g., ECB supervisory data), raising questions about whether confounders were designed to be solvable.
  • Confounders are explicitly and artificially injected; real financial networks have latent, ill-defined confounders not captured by a regional identifier.
  • The counterfactual generation runtime scales super-linearly (18.4s for \(N=2000\)), which may be limiting for real-time stress scenarios with larger networks.
  • No temporal modeling: the framework treats the financial network as static; financial contagion is inherently dynamic.
  • The adversarial causal regularization requires specifying confounder identity \(C\) at training time — a strong assumption in practice.

Relevance to Agentic AI / LLM Agents

CausalGraphX is not directly about LLM agents, but it is methodologically relevant to the causal reasoning and explainability strand of agentic AI research. Agentic systems increasingly need to generate causal "what-if" explanations and minimal intervention plans — precisely the capability CausalGraphX operationalizes via counterfactual optimization over structured graphs. The adversarial confounder-invariance technique (minimax causal regularization) is directly applicable to agent world models that must distinguish genuine cause-effect relationships from spurious distributional patterns. For agents operating in high-stakes domains (finance, healthcare, infrastructure), the counterfactual explanation generator provides a template for generating actionable, sparse intervention recommendations grounded in learned causal structure rather than post-hoc correlation attribution.