How your supporters and opponents define your interestingness

Abstract

How can one determine whether a data mining method extracts interesting patterns? The paper deals with this core question in the context of unsupervised problems with binary data. We formalize the quality of a data mining method by identifying patterns – the supporters and opponents – which are related to a pattern extracted by a method. We define a typology offering a global picture of the methods based on two complementary criteria to evaluate and interpret their interests. The quality of a data mining method is quantified via an evaluation complexity analysis based on the number of supporters and opponents of a pattern extracted by the method. We provide an experimental study on the evaluation of the quality of the methods.