GEC: An Evolutionary Approach for Evolving Classifiers

Using an evolutionary approach for evolving classifiers can simplify the classification task. It requires no domain knowledge of the data to be classified nor the requirement to decide which attribute to select for partitioning. Our method, called the Genetic Evolved Classifier (GEC), uses a simple structured genetic algorithm to evolve classifiers. Besides being able to evolve rules to classify data in to multi-classes, it also provides a simple way to partition continuous data into discrete intervals, i.e., transform all types of attribute values into enumerable types. Experiment results shows that our approach is comparable to methods like C4.5, Fuzzy-ID3 (F-ID3), and probabilistic models such as modified Naïve-Bayesian classifiers. For the adult census database, our average result is around 3% worse compared to C4.5. For the iris database, our method performs worst, 6% worse than F-ID3 and 3% worse than C4.5. But for the yeast database, our method outperformed past method (probabilistic method) around 7%. Finally, for the wine database, our method is most superior. Our method outperformed C4.5 and F-ID3 by 5% and 3% respectively. This proves that our method is comparable with other method and produces promising results.