Diversification for better classification trees

Classification trees are widely used in the data mining community. Typically, trees are constructed to try and maximize their mean classification accuracy. In this paper, we propose an alternative to using the mean accuracy as the performance measure of a tree. We investigate the use of various percentiles (representing the risk aversion of a decision maker) of the distribution of classification accuracy in place of the mean. We develop a genetic algorithm (GA) to build decision trees based on this new criterion. We develop this GA further by explicitly creating diversity in the population by simultaneously considering two fitness criteria within the GA. We show that our bicriterion GA performs quite well, scales up to handle large data sets, and requires a small sample of the original data to build a good decision tree.