Counterfactual linguistic rule-based explanations based on locally relevant causal mechanisms
Te Zhang, Christian Wagner · 2025
Counterfactual (CF) explanations provide a potentially powerful mechanism to deliver meaningful explanations of AI decisions. CF explanations are convincing when they reflect causal relationships between variables, because humans are cause-effect thinkers. Prior work has established a rule generation framework called CF-MABLAR, which is designed to generate causal rules that provide CF explanations. However, in the real-world, an effect is often the result of multiple causal mechanisms, and rules obtained by CF-MABLAR may not capture the actual causal mechanism that leads to the effect, which we called the locally relevant causal mechanism. Consequently, CF explanations generated by CF-MABLAR have the risk of containing redundant components, which reduces the explainability of the obtained CF explanations. To address this issue, in this paper, we provide a detailed discussion about two key aspects of generating CF explanations from a causal perspective: 1) which variables require intervention and 2) what magnitude of an intervention is needed. We propose CF-MABLAR-local which allows users to generate CF explanations based on locally relevant causal mechanisms. We conduct experiments on several real-world data sets to compare CF explanations generated through different methods, and analyse the impact of different parameterizations in CF-MABLAR-local.