Simplest Rules Characterizing Classes Generated by δ-Free Sets

摘要

We present a new approach that provides the simplest rules characterizing classes with respect to their left-hand sides. This approach is based on a condensed representation (δ-free sets) of data which is efficiently computed. Produced rules have a minimal body (i.e. any subset of the left-hand side of a rule does not enable to conclude on the same class value). We show a sensible sufficient condition that avoids important classification conflicts. Experiments show that the number of rules characterizing classes drastically decreases. The technique is operational for large data sets and can be used even in the difficult context of highly-correlated data where other algorithms fail.