A0356
Title: Classification method for corrupted label data using density ratio
Authors: Masaaki Okabe - Doshisha University (Japan) [presenting]
Hiroshi Yadohisa - Doshisha University (Japan)
Abstract: During the training of the classification model, the labels of the training data were assumed to be correct. However, human errors such as mislabeling can result in incorrect labels being assigned to objects. If training data contain incorrect labels, the classification accuracy of the model may be reduced. A prior study proposed a classification method that uses the balanced error rate as the objective function for training from corrupted label data. This method assumes independence between features and corrupted occurrences, given a true label, and specifically addresses cases where mislabeling occurs randomly. However, classification may not work well when label corruptness is correlated with features. If label errors depend on the features, they are correlated with the features. The focus is on the arrangement of Menon's assumptions, proposing an estimation of classification models with relaxed assumptions compared to existing methods by using the ratio of the contaminated label distribution to the true label distribution of the obtained data.
