The Analysis of Explainable AI via Notion of Congruence

Explainable AI has become a popular topic which draw attentions of many researchers, and since the inception of this field, the most popular method of model explanation is undoubtedly using Decision Tree to approximate black box model, due to its straightforwardness and ease of explanation. However, Decision Tree is single dimensional model without any nuance in the decision making which may be intrinsic to the structural properties of the black box model itself. With this, we propose using a structured framework in the form of Probabilistic Assumption-based Argumentation (PABA) framework in place of Decision Tree which allows for more complex relationship between arguments within the model, and demonstrate an application using PABA framework as an alternative to Decision Tree by establishing “notion of congruence” between an example black box model and its structured argumentation framework (PABA) predictor counterpart, that is, some measurable similarity between both models along with establishing methodology to evaluate their relative faithfulness. In order to accomplish that, we partially translated both models to PABA framework first, then extract the explanations from both models using an adaptation of two existing techniques (Anchor and LORE), which is then successfully compared using established notion of congruence to get a numerical measurement of faithfulness. By proving that this is possible, our study suggests that improvement on the framework level of machine explanation problem is a possibility.