A Framework of the Semi-supervised Multi-label Classification with Non-uniformly Distributed Incomplete Labels

Abstract

In real world applications, the problem of incomplete labels is frequently encountered. These incomplete labels decrease the accuracy of the supervised classification model because of a lack of negative examples and the non-uniform distribution of the missing labels. In this paper, we propose a framework of the semi-supervised multi-label classification which can learn with the incompletely labeled training data, especially for the missing labels whose distribution is not a uniform distribution. With a modified instance weighted k nearest neighbor classifier, this framework recovers the labels of the training data, including both the incomplete labeled part and the unlabeled part, by iteratively updating the weight of each training instance in an acceptable execution time. The experimental results verify that the classification model trained from the recovered training data generates better prediction results in the testing phase.