Small sample learning algorithm based on novel hybrid class-labeling technique

A small sample learning algorithm based on a novel hybrid class-labeling technique (HCLT) was proposed in order to address the learning problem resulting from the underrepresented labeled training set in computer-aided diagnosis(CAD). The abundant unlabeled samples were labeled by HCLT with three diverse class labeling schemes respectively from the view point of geometric similarity, probabilistic distribution and semantic concept. Only those unlabeled samples which get the unanimous labeling results from three different labeling schemes were added to the training set in order to enlarge the labeled training set. The memberships of pseudo-labeled samples were introduced to fuzzy support vector machine (FSVM) in order to reduce the adverse effects for learning performance resulting from the still existing labeling mistakes. The contributions of pseudo-labeled samples to learning task were determined by their memberships. Classification experiment results based on datasets in UCI show that the proposed algorithm can deal with the small sample learning problem. The algorithm has less mistakes and better classification performance compared with the other algorithms which adopt the single labeling scheme. © 2016, Zhejiang University. All right reserved.