A Semi-Supervised Approach Based on k-Nearest Neighbor

Abstract

k-nearest neighbor (k-NN) needs sufficient labeled instances for training in order to get robust performance; otherwise, performance deterioration occurs if training sets are small. In this paper, a novel algorithm, namely ordinal semi-supervised k-NN, is proposed to handle cases with a few labeled training instances. This algorithm consists of two parts: instance ranking and semi-supervised learning (SSL). Using SSL, the performance of k-NN with small training sets can be improved because SSL enlarges the training set by including unlabeled instances with their predicted labels. Instance ranking is used to pick up the unlabeled instances that are included to the training set. It gives priority to the unlabeled instances that are closer to class boundaries because they are more likely to be correctly predicted (these instances are called high confidence prediction instances). Thus, SSL benefits from high confidence prediction instances if they are added into the training set early. Experimental results demonstrate that the proposed algorithm outperforms the semi-supervised k-NN and the conventional k-NN algorithms if the training set is small.&nbsp

MC‐KV: A Prognosis‐Oriented Classifier Based on Semi‐Supervised Learning for Molecular Subtyping of Colorectal Cancer

“…Here we proposed a new semi-supervised algorithm with repeated randomized sampling, in which the follow-up data were separated into 1-, 3-, and 5-years and then integrated into one classifier. Among various semi-supervised classification algorithms, KNN is always favored for its high stability and efficiency [27,28] and was adopted as the core algorithm in our classification model. Additionally, we utilized the prior knowledge-based pathway information instead of single gene expression data as input data, to ensure a high information integration rate and biological credibility.…”