Explaining Causal Models with Argumentation: the Case of Bi-variate Reinforcement

Causal models can be translated into argumentation frameworks to make AI decisions more explainable. The paper proposes a new method that converts causal models—which show how different factors influence outcomes—into structured arguments that humans can more easily understand and evaluate.

Argumentation frameworks offer a principled way to explain AI outputs by mimicking how humans naturally reason about competing factors. By reinterpreting properties of argumentation semantics as "explanation moulds," the researchers create a systematic approach to characterizing relationships in causal models in argumentative terms, rather than relying on informal explanations.

The "bi-variate reinforcement" property enables explanations that capture how two factors jointly support or oppose conclusions. This approach allows the framework to explain decisions based on multiple interacting causes, which is more realistic than explanations that only consider single factors in isolation.

The methodology was theoretically validated against both explanatory and argumentative quality standards. The authors tested whether their argumentative explanations satisfied desirable properties—suggesting this approach produces explanations that are both logically sound and genuinely informative about how the causal model reached its conclusions.

This work bridges formal argumentation theory and explainable AI, offering a structured alternative to black-box explanation methods. Rather than post-hoc explanations added after a model decides, this approach builds explanations directly from the causal reasoning underlying the model's decision.