Mislabeled examples affect the performance of supervised learning algorithms. Two novel approaches to this problem are presented in this Thesis. Both methods build on the hypothesis that the large margin and the soft margin principles of support vector machines provide the characteristics to select mislabeled examples. Extensive experimental results on several datasets support this hypothesis. The support vectors of the one-class and two-class SVM classifiers captures around 85% and 99% of the randomly generated label noise examples (10% of the training data) on two character recognition datasets. The numbers of examples that need to be reviewed can be reduced by creating a two-class SVM classifier with the non-support vector examples, and then by only reviewing the support vector examples based on their classification score from the classifier. Experimental results on four datasets show that this method removes around 95% of the mislabeled examples by reviewing only around about 14% of the training data. The parameter independence of this method is also verified through the experiments. All the experimental results show that most of the label noise examples can be removed by (re-)examining the selective support vector examples. This property can be very useful while building large labeled datasets.

Large scale datasets collected using non-expert labelers are prone to labeling errors. Errors in the given labels or label noise affect the classifier performance, classifier complexity, class proportions, etc. It may be that a relatively small, but important class needs to have all its examples identified. Typical solutions to the label noise problem involve creating classifiers that are robust or tolerant to errors in the labels, or removing the suspected examples using machine learning algorithms. Finding the label noise examples through a manual review process is largely unexplored due to the cost and time factors involved. Nevertheless, we believe it is the only way to create a label noise free dataset. This dissertation proposes a solution exploiting the characteristics of the Support Vector Machine (SVM) classifier and the sparsity of its solution representation to identify uniform random label noise examples in a dataset. Application of this method is illustrated with problems involving two real-world large scale datasets. This dissertation also presents results for datasets that contain adversarial label noise. A simple extension of this method to a semi-supervised learning approach is also presented. The results show that most mislabels are quickly and effectively identified by the approaches developed in this dissertation.