Weighted Least Squares Support Vector Machine for Semi-supervised Classification

The recently proposed semi-supervised least squares support vector machine (SLS-SVM), extends support vector machine (SVM) to semi-supervised learning field. However, the support value in SLS-SVM is not zero and the solution is lack of sparseness. To overcome this drawback, a weighted semi-supervised SLS-SVM (WSLS-SVM) is proposed in this paper, where the impact of labeled and unlabeled samples can be controlled by weighting the corresponding error. It is basically a pruning method according to the sorted weight of estimation error. To solve the proposed classifier, an efficient progressive learning algorithm is presented to reduce the iteration. Experimental results on several benchmarks data sets confirm the sparseness and the effectiveness of the proposed method.