This letter provides a simple extension of boosting methods for binary data where the probability of mislabeling depends on the label of an example. Loss functions are derived from the statistical perspective, which is based on likelihood analysis. Our proposed methods can be interpreted as a correction of the decision boundary of observed labels. This interpretation partially relates to cost-sensitive learning, a classification method for the case in which the ratio of two labels in a dataset is skewed. Numerical experiments show that the proposed methods work well for asymmetric mislabeled data even when the probabilities of mislabeling may not be precisely specified.
- Asymmetric mislabeling mechanism
- Bayes error rate
ASJC Scopus subject areas
- Statistics and Probability
- Statistics, Probability and Uncertainty