Similarity-evaluation-based evolving of flexible neural trees for imbalanced classification

Abstract

A flexible neural tree (FNT) is a special and effective kind of artificial neural network that can search optimal network structures using tree structure evolving algorithms, leading to its high performance in many real classification and prediction problems. However, the standard FNT model does not take into account tree structure similarities in its evolving process. This may result to a rapid and significant decrease of its tree structure population. Standard FNT also suffers from imbalanced data. In this study, we propose a new similarity evaluation method for FNT (SEFNT) to keep the population diversity and deal with the imbalanced data. The main idea of SEFNT is twofold. First, the difference between two nodes is introduced for similarity measurement. Second, the node position and height of trees are also taken into account for accurate tree structure distance measurement. SEFNT uses imbalanced fitness function to control its evolving procedure to deal with imbalanced problems. We compare SEFNT with 10 imbalanced methods on 46 KEEL datasets and 10 UCI datasets. The experimental results show that our approach can significantly improve the classification performance of FNT. In the comparisons with other algorithms, SEFNT shows significantly better performance. We also apply the proposed method to a practical imbalanced classification problem, that is, Internet video traffic identification. The results imply that our method is effective in dealing with practical problems.