Evaluating Restrictions in Pattern Based Classifiers

Generalizations, also known as patterns, are in the core of many learning systems. A key component for automatically mine generalizations is to define the predicate to select the most important ones. This predicate is usually expressed as a conjunction of restrictions. In this paper, we present an experimental study of some of the most used restrictions: the minimal support threshold, the jumping pattern and the minimal pattern. This study that uses 93 databases and two different classifiers reveals some interesting results, including one that should be very useful for building better classifiers: using minimal patterns could degrade the accuracy.