Explainable Machine Learning via Argumentation

This paper presents a general Explainable Machine Learning framework and methodology based on Argumentation (ArgEML). The flexible reasoning form of argumentation in the face of unknown and incomplete information together with the direct link of argumentation to justification and explanation enables the development of a natural form of explainablemachine learning. In this form of learning the explanations are useful not only for supporting the final predictions but also play a significant role in the learning process itself. The paper defines the basic theoretical notions of ArgEML together with its main machine learning operators and method of application. It describes how such an argumentation-based approach can give a flexible way for learning that recognizes difficult cases (with respect to the current available training data) and separates these cases out not as definite predictive cases but as cases where it is more appropriate to explainably analyze the alternative predictions. Using the argumentation-based explanations we can partition the problem space into groups characterized by the basic argumentative tension between arguments for and against the alternatives. The paper presents a first evaluation of the approach by applying the ArgEML learning methodology both on artificial and on real-life datasets.