User-Centred Argumentation Analysis of Local Explanations in Explainable AI

A local explanation method (LE) is basically a two-step procedure: first construct an interpretable model to substitute the given black-box model in need of explanations; then extract an explanation from the constructed substitution model and present it to the user as if it were genuinely extracted from the black-box model. Though serving well the didatic purpose, whether such an explanation method sustains any user’s real belief is unclear. This paper starts with a stance that an expert user does not take what a LE presents literally, but engages in a quite intricate “post-hoc” internal reasoning using analogical arguments to selectively transfer some properties of the presented explanation (that LE extracted from the substitution model) to the explanation of the actually queried model. We then reconstruct the structures “reason  therefore  conclusion” of these analogical arguments and study conditions for ensuring the truth of reason; as well as conditions for ensuring that conclusion follows deductively from reason. It is argued that the presented theoretical findings shed light on how to simulate the whole post-hoc reasoning of a hypothetical expert user after a dialogical interaction with a local explanation method. Broadly speaking, the paper suggests a promising direction to extend system-centred XAI to user-centred XAI.