Classifying Data

Modern day classifiers achieve impressive performance but almost always suffer from the lack of transparency regarding the classifications. Even when the classifications are accompanied with e.g. instantiated parameters and/or weights obtained while training a neural network, that additional information is hardly human-interpretable. We seek to overcome the transparency problem by using argumentation-based approaches to classification.

Argumentation methodologies allow us to represent and reason with classification data via directed graphs. In the AA-CBR methodology, the graphs have as nodes the labelled data points and the edges represent conflicts among them according to their features and classes. Given a new data point, argumentation drives the resolution of conflicts and provides a classification. This is then supplemented by a sub-graph representing the relevant data points and their relationships. Such sub-graphs yield dialectical as well as rule-based explanations for the classification, relying on the argumentative structure of the data.

Papers