Search for better random forests with an tree selection method

Random forest is an ensemble method with high classification performance by voting the results of individual tree classifiers. However, owing to the complexity of data distribution in high dimensional space, a random forest may include bad trees that can result in wrong results. As a consequence, inappropriate ensemble classification decision will be made if there are a large proportion of bad trees included in a random forest. In this paper, we propose a tree selection method which aims to optimize the tree selection process so that only good trees are selected and included in a random forest. Experimental results on both the UCI and real world datasets have demonstrated that the proposed method could generate a random forest with higher performance with regard to the classification accuracy and the error bound than the random forests generated by Breiman's method.