Towards Argumentative Decision Graphs: Learning Argumentation Graphs from Data

In this paper we present a novel data-mining model called argumentative decision graphs (ADG). An ADG is a special argumentation framework where arguments have a rule-based structure and an attack relation is defined among arguments. ADGs are graph-like models learnt from data in a supervised way that can be used for classification tasks. As in a decision tree, given a set of input features, an ADG returns the value of the target variable. Unlike decision trees, the output of an ADG can be also an undecided status, occurring when the model does not have enough reasons to predict a value for the target variable. This is due to the use of argumentation semantics to identify what arguments of an ADG are accepted and consequently make a prediction about the target variable. Unlike Bayesian Networks, ADGs are not required to be acyclic, but they can have any topology. Advantages of ADGs are the possibility of using different semantics to make predictions, the ability to deal with incomplete input data and to generate compact explanations. We evaluate a preliminary greedy algorithm to learn an ADG from data using public datasets and we compare our results with Decision Tree in terms of balanced accuracy and size of the model. Our results provide evidence to further progress our research.