Ordinal-imbalanced data classification through data reduction by singular value decomposing truncation

Abstract

In the real world, multi-class ordinal data classification problems occur frequently. Most ordinal classifiers were constructed assuming that a class distribution is balanced, but most ordinal data have skewed class distributions. The class imbalance degrades the performance of traditional learning. Many papers address the difficulty of the class imbalance but pay little attention to the imbalance arising in ordinary class data. So, we analyze the imbalance issue of ordinal data. This paper introduces a matrix factorization method of preprocessing algorithm called singular value decomposition (SVD) truncation for ordinal classification. It has a role of noise reduction which is an effective method for the imbalance issue. Also, the proposed method diminishes an overlapping area that also has a positive effect on dealing the imbalance. Furthermore, the suggested preprocess algorithm does not modify class distributions. It complements the weaknesses of existing sampling methods such as loss information and over-fitting. We used the wkNN algorithm for ordinal classification after the proposed preprocessing technique for experiments. Experimental results on actual ordinal data verify the usefulness of the methodology. © 2019 IISE Annual Conference and Expo 2019. All rights reserved.