Missing Value Estimation Using a Sub-bagging Model of Self-organizing Maps

Abstract

Missing value estimation is an important task in data mining and analysis of data containing missing values. The purpose of this study is to improve the accuracy of missing value estimation using self-organizing maps (SOMs), which have been studied in recent years. We have focused on the ensemble learning algorithms based on bootstrap sampling that have been successfully used in recent years, in cluster ensembles and pattern recognition. In the present study, in order to improve the accuracy of missing values estimation, we applied the bagging and sub-bagging major ensemble learning algorithms to SOM. We tested the effectiveness of the proposed methods through computational experiments using bench mark data sets published in the UCI Machine Learning Repository. The reproducibility error with respect to the artificial missing values was evaluated. The experimental results show that our methods were better in estimation using conventional SOM and simple ensemble of SOMs, from the viewpoint of the accuracy of missing value estimation. Further, sub-bagging was confirmed to tend to have higher accuracy than bagging