Sample Selection And Fuzzy Rough Set Based Soft Margin Support Vector Machine

For most supervised learning approaches, instances in the training set are often selectedrandomly from a pool of unlabeled instances and annotated by expert. But labels are oftenexpensive to get. If we use active learning approach to select informative instances tocomprise training set, then the necessity of labeled instances can be eliminated; Supportvector machine has excellent abilities in comparison with other supervised learningapproaches. Combining support vector machine with active learning, a new sample selectionapproach is proposed in this paper. For each instance in an unlabeled pool and each possiblelabel, the distance to the hyperplane is considered when this instance is put into the trainingset to learn an hyperplane. Our approach selects such an instance which is the nearest to thenew hyperplane in the worst situation of labels. The informativeness of the instances can beassured in constructing new hyperplane to a great extent, so a classifier with higher predictioncan be learned on small sized instances.As inconsistency between conditional attributes and labels of instances is not consideredwhen we are learning an optimal hyperplane by applying support vector machine approach,the optimal hyperplane may be sensitive to noises; the concept of inconsistency is definedexplicitly in rough set, in which a dependency function is used to measure the consistentdegree between conditional attributes and decision attributes of training instances. Rough setis an effective tool in dealing with problems of fuzziness and uncertainty. Fuzzy rough setbased soft margin support vector machine is proposed in this paper. It considers inconsistencybetween conditional attributes and decision labels of instances in the training set, allowsinstances to be misclassified during constructing the optimal hyperplane in the trainingprocess, and learns an optimal hyperplane by considering both maximal margin and minimalmisclassification errors. So our approach is less sensitive to noises. Experimental results showthe efficiency of the approach.