A Clustering-Inspired Quality Measure for Exceptional Preferences Mining - Design Choices and Consequences

Ruben Franciscus Adrianus Verhaegh, Jacco Johannes Egbert Kiezebrink, Frank Nusteling, Arnaud Wander André Rio, Márton Bendegúz Bendicsek, Wouter Duivesteijn, Rianne Margaretha Schouten

Abstract: Exceptional Preferences Mining (EPM) combines the research fields of Preference Learning and Exceptional Model Mining. It is a local pattern mining task, where we try to find coherent subgroups of the dataset featuring unusual preferences between a fixed set of labels. We introduce a new quality measure for Exceptional Preferences Mining, inspired by concepts from Clustering. On top of that, we draw conclusions on two design choices that must necessarily be made whenever one defines a quality measure for any version of Exceptional Model Mining: on the one hand, exceptional behavior is easily (spuriously) found in tiny subgroups, so what is the best way to compensate for that; on the other hand, when gauging exceptionality of a subgroup’s behavior, what does one use as reference for the normal behavior? We find that the choice of correction factor not only influences the subgroup size but it also effects the presumed exceptionality of found subgroups. The entropy function allows for detecting exceptional subgroups of a meaningful size, both when a candidate subgroup is evaluated against its complement and against the entire dataset.