DEMST-KNN: A Novel Classification Framework to Solve Imbalanced Multi-class Problem

Abstract

Imbalanced multi-classification problem is an important research hotspot in machine learning. For multiple categories of data, an important method is divide-and-conquer, which covers the decomposition strategy and ensemble rule. However, with the increase of category, the number of poor binary-classifiers increases, which greatly reduces the classification accuracy. In this paper, a novel classification framework to solve current problem of multi-classification method for imbalanced data was proposed, which including a decomposition strategy based on maximum spanning tree, an ensemble rule based on node degree to find the optimal combination of binary-classifiers and a specific KNN algorithm which can dynamically adjust K neighborhood to increase probability of minority class. Finally we experimented on nine public data sets in different fields and compared our proposed classification framework with the most popular framework. The experimental results show that the proposed classification framework can greatly improve the classification accuracy of minority class.