Abstract

We describe PADUA, a protocol designed to support two agents debating a classification by offering arguments based on association rules mined from individual datasets. We motivate the style of argumentation supported by PADUA, and describe the protocol. We discuss the strategies and tactics that can be employed by agents participating in a PADUA dialogue. PADUA is applied to a typical problem in the classification of routine claims for a hypothetical welfare benefit. We particularly address the problems that arise from the extensive number of misclassified examples typically found in such domains, where the high error rate is a widely recognised problem. We give examples of the use of PADUA in this domain, and explore in particular the effect of intermediate predicates. We have also done a large scale evaluation designed to test the effectiveness of using PADUA to detect misclassified examples, and to provide a comparison with other classification systems.