Label Estimation Method with Modifications for Unreliable Examples in Taming

Abstract

Methods for improving learning accuracy by utilizing a plurality of data sets with different reliabilities have been studied extensively. Unreliable data sets often include data with incorrect labels, and the accuracy of learning from such data sets is thus affected. Here, we focused on a learning problem, Taming, which deals with two kinds of data sets with different reliabilities. We propose a label estimation method for use in data sets that include data with incorrect labels. The proposed method is an extension of BaggTaming, which has been proposed as a solution to Taming. We conducted experiments to verify the effectiveness of the proposed method by using a benchmark data set in which the labels were intentionally changed to make them incorrect. We confirmed that learning accuracy could be improved by using the proposed method and data sets with modified labels.