A semi-supervised k-nearest neighbor algorithm based on data editing

YATSI may suffer more from the common problem in semi-supervised learning, i.e. the performance is badly influenced due to the unlabeled examples may often be wrongly labeled. In this paper a semi-supervised k-nearest neighbor classifier named De-YATSI (YATSI with Data Editing) is proposed. A data editing based on estimating class conditional probability is used to identify and discard mislabeled examples of the pre-labeled data set. A k-nearest neighbor classifier with weights is trained by the labeled data set and the edited “pre-labeled” data set. Experiments on UCI datasets show that DE-YATSI could more effectively and stably utilize the unlabeled examples to improve classification accuracy than YATSI.