Computing Actual Causes for Neural Network Predictions under Structured Causal Inputs

Explaining neural network (NN) predictions is critical for trustworthy AI, yet most existing methods treat input features as independent. This assumption often leads to misleading or spurious explanations when real-world data contains structured causal dependencies. The authors address this by formalizing explanations as Halpern-Pearl (HP) actual causes, which identify minimal sets of features that, if changed, would alter a prediction. To ensure these explanations respect the underlying causal structure, the authors model input dependencies using Boolean Structural Causal Models (SCMs).

To compute these causes, the authors introduce the CausEx-BaB algorithm. Instead of brute-force enumeration, which is computationally infeasible for large feature sets, the algorithm uses interval bound propagation (IBP) and branch-and-bound techniques. By representing potential causal interventions and contingencies as regions, the algorithm can prune large parts of the search space that are guaranteed not to contain valid causes, while recursively refining inconclusive regions.

The authors demonstrate that CausEx-BaB is both sound and complete, providing formal guarantees for the minimality of the identified causes. In experiments using synthetic SCM-NN benchmarks, the algorithm significantly outperforms brute-force, integer linear programming (ILP), and heuristic search baselines. It successfully handles SCMs with up to 28 nodes and search spaces containing over 10^13 candidate pairs within a 180-second budget.

A case study on the U.S. Supplemental Nutrition Assistance Program (SNAP) highlights the practical importance of this approach. When input dependencies are ignored, the number of reported causes is inflated, with 14.9% of identified causes being spurious under the defined SCM. This underscores that failing to account for causal structure can lead to fundamentally incorrect explanations for automated decisions.

This research provides a rigorous framework for debugging and auditing neural networks in domains where features are not independent, such as finance, healthcare, and public policy. By moving beyond simple feature attribution to structural causal reasoning, the authors enable developers to identify the actual root causes of model predictions, thereby improving the reliability and transparency of AI systems in high-stakes environments.