A Transductive Support Vector Machine Algorithm Based on Spectral Clustering
Xu Yu, Jing Yang, Jianpei Zhang · AASRI Procedia · 2012
As transductive inference makes full use of the distribution information of unlabeled samples, so compared with traditional inductive inference, it is always more precise. For the question that transductive support vector machine algorithm (TSVM) is with a low classification precision and an unstable learning performance, in this paper, a transductive support vector machine algorithm based on spectral clustering (TSVMSC) is proposed. Firstly, the algorithm utilizes spectral clustering algorithm to divide the unlabeled samples into several clusters, then labels them with the same class, finally makes transductive inference on the mixed data set composed by both labeled and unlabeled samples. It is not necessary to estimate the ration of the positive samples to the negative samples, so the stability can be improved largely. Meanwhile, the prior distribution information of the unlabeled samples can further be extracted, so the classification precision can be improved effectively. The experiments show that TSVMSC exhibits a superiority over TSVM both at the stable performance and classification precision.