Proposed a general noise added model that addresses the classification problem in which noise is included in the label. The purpose is to express Semi-supervised learning, PU learning (ground truth + without label), and problems with outliers unifiedly.
Proposed classification algorithm using EM algorithm and variational Bayesian method.
Others
Expressed these problems (semi-Supervised, PU, contains outlier) that sightly seemed the different problems with the same matrix (M x L); M (true label: positive, negative, outlier) & L (obtained label: positive, negative, no-labeled). In the matrix, p expressed the probability of label flips is contains.
In the experiments, the simulation are implemented under some conditions of p and data size.
Basic Information
Link
https://ci.nii.ac.jp/naid/40021403047
Overview
Others
Reference (for understanding)