Sensitivity analysis on initial classifier accuracy in fuzziness based semi-supervised learning

Abstract Semi-supervised learning can be described from different perspectives, which plays a crucial role in the study of machine learning. In this study, a new aspect of semi-supervised learning is explored by investigating the divide-and-conquer strategy based on fuzziness to improve the performance of classifiers. In such an approach, adding a category of samples with low fuzziness in the training set can improve the training accuracy, which is experimentally confirmed and explained in the theory of learning from noisy data. The significance of initial accuracy of a base classifier in improving classifier’s performance is further studied. It is observed that the initial accuracy of a base classifier has a significant impact on the improvement of classifier’s performance. Experimental results exhibit that the improvement of accuracy, which is sensitive to the base classifier, attains its maximum when the initial accuracy is between 70% and 80%.