Evolutionary Algorithms For Effective Classification

Abstract

The amount of raw data being accumulated in the databases is increasing at an incon- ceivable rate. However, these data-rich databases are poor in providing substantial information. This is where data mining comes into picture. Speci cally, data mining is \the process of extracting or mining information from large amount of data". Data classi cation has been an active area of research in data mining. It consists of assign- ing a data instance to one of the prede ned classes/groups based upon the knowledge gained from previously seen (classi ed) data. Real world problems demand classi ers that are accurate as well as easy to inter- pret. Traditional comprehensible classi ers like decision trees are very accurate at classifying new instances, but with increase in the size of datasets and/or the number of classes, and/or with increase in number of attributes, the trees induced are very large in size and hence difficult to interpret. On the other hand, evolutionary algo- rithms like Genetic Programming(GP) when applied to classi cation problems give trees that have smaller size but are not very accurate. Thus the work proposes an algorithm GPeCT, that employs GP as an optimiza- tion technique to yield efficient and e ffective classifi cation. The goal is to evolve a classi er that performs a trade-o between accuracy and comprehensibility in order to produce an optimal decision tree classi er (for n-class where n>=2) using GP. When evaluated on some benchmark datasets, the proposed algorithm obtained by merging Genetic Programming and decision tree outperforms the traditional classi fi- cation techniques in terms of a combination of accuracy and comprehensibility.