Using Support Vector Machine for Emotion Classification in presence of Noise Label
MetadataShow full item record
The ability to deal with noise label is very important for machine learning classification algorithms. For some kinds of data, since there exist ambiguity and uncertainty in the data label, we will have to consider those noise in our learning algorithm to handle with this issue. In this thesis, the support vector machines (SVMs) has been performed as the classifier to classify the data and Biggio proposed one Label Noise robust SVM (LN-robust SVM) to improve the robustness of SVMs by simply correcting the kernel matrix with another structure when solving SVM optimization problem. However, the effectiveness of LN-robust SVM may not work in all cases of the data. In some cases, the LN-robust SVM even reduce the accuracy compared with the standard SVM classification. This problem becomes important when it comes to multiple classes classification scenario since we need to find an effective criteria to determine whether to apply LN-robust SVM or not in each binary classification. To answer this question, this thesis conduct a comprehensive investigation on the effects of the data separability on the performance of LN-robust SVMs. We have shown that the data separability is a critical criteria that will affect the effectiveness of LN-robust SVMs a lot. Based on this criteria, we then proposed a framework for applying LN-roust SVMs into multiple classes classification developed for human emotion classification. The accuracy increases from 81% to 85% after applying the framework.