Symbolic learning supervised algorithms are often used in Knowledge Discovery in Database. In most cases, the knowledge induced by these algorithms can be expressed as a set of rules. Those rules, however, have a bias that is related to the support and accuracy inherent to the induced classifier. Thus, although the set of rules that constitute the classifier may represent the implicit knowledge of the database, every rule in this set tends to neglect certain desirable properties, such as novelty, sensitivity, and others, for these properties are much more related to each individual rule than to the entire classifier. This project proposes the use of Evolutionary Computation techniques to build new rules with specific properties from rules induced by learning algorithms, and/or other knowledge rules. Thus, it is necessary to design and develop a data structure capable of representing knowledge rules in such a way it will be possible to apply on this structure different methods used by Evolutionary Algorithms to create new rules. Several Evolutionary Algorithm methods will be implemented in Perl as a module for the generation of knowledge rules with specific properties, and will be evaluated using natural and real world databases. This new module will be built into the computation environment for knowledge extraction we are developing in our research lab.
News published in Agência FAPESP Newsletter about the scholarship: