Emerging Pattern Based Classification in Relational Data Mining

The usage of descriptive data mining methods for predictive purposes is a recent trend in data mining research. It is well motivated by the understandability of learned models, the limitation of the so-called "horizon effect" and by the fact that it is a multi-task solution. In particular, associative classification, whose main idea is to exploit association rules discovery approaches in classification, gathered a lot of attention in recent years. A similar idea is represented by the use of emerging patterns discovery for classification purposes. Emerging Patterns are classes of regularities whose support significantly changes from one class to another and the main idea is to exploit class characterization provided by discovered emerging patterns for class labeling. In this paper we propose and compare two distinct emerging patterns based classification approaches that work in the relational setting. Experiments empirically prove the effectiveness of both approaches and confirm the advantage with respect to associative classification.