Efficient implementation and experimental testing of transductive algorithms for predicting with confidence
Support Vector Machines (SVM's) and other kernel based methods have grown in popularity in recent years. Although they have many benefits, such as the ability to deal with a large number of parameters, one drawback of these successful techniques is their lack of the ability to provide rigorous confidence measures for the predictions they make. This thesis is devoted to the efficient implementation and experimental testing of transductive algorithms developed at the computer science department, Royal Holloway. The algorithms are tested against several benchmark data sets, and methods for comparing quantitative confidence values are described and evaluated. These techniques and other machine-learning methods are also applied to the industrial application of fault diagnosis and automated repair. An extensive case study of applying these machine learning techniques to a real-world problem is carried out. Many problems such as data collection and representation -- which are common to most real-world applications of machine learning techniques, but sometimes over-sighted in literature -- are highlighted and discussed.