Neural Network Explanations via Halpern-Pearl Actual Causes

A recent paper on arXiv (2608.03772) tackles the issue of elucidating neural network predictions when the input features exhibit dependencies. The researchers define explanations in terms of Halpern-Pearl (HP) actual causes, utilizing Boolean Structural Causal Models (SCMs) to represent these input dependencies. They employ bound propagation and branch-and-bound methods to derive HP causes, ensuring formal completeness and minimality. Their experiments demonstrate significant scalability enhancements compared to brute-force and ILP baselines, as well as superior performance against heuristic searches as graph sizes increase. The approach identifies all minimal actual causes in search spaces reaching up to 2.3×10^13 candidate (cause, contingency) pairs on SCMs with a maximum of 28 nodes in a timely manner. This research aids in developing trustworthy AI by providing more precise explanations for structured inputs.