SEQUENTIAL CONSTRUCTION OF FEATURES BASED ON GENETICALLY TRANSFORMED DATA
Exploration of real data sets is a complex task that often involves tiresome, manual parameter tuning. Such manual operation, aimed at transformations of data that enable discovery of interesting patterns, only rarely guarantees any thorough examination of all promising combinations of parameter values. To avoid this inconvenience, we present a universal data transformation approach that has the ability to conduct fully automatic adjustments of parameter values. The main mechanism is based on a genetic algorithm designed to search for parameter settings that are optimal with respect to a pre-defined objective function. As an illustration of the procedure we present a system that improves classification of vowels by constructive induction of new features (attributes). The new features are created in a process that is entirely automatic: the original data are transformed with a set of sequentially applied operators, the parameters of which are incorporated in a genome and thus easily controlled by the genetic search engine. The results of several conducted experiments prove the usefulness of the proposed approach.