New approach to probabilistic classification based on gaussian process and support vector machine

Abstract

In recent years, Support Vector Machine (SVM) has achieved the feat of classifying a large amount of data more efficiently and accurately compared to most of the existing data classifying algorithms. Gaussian Process Classification (GPC), on the other hand, has the advantage of providing probabilistic information on the classification but atthe cost of high computational requirements. In this thesis,two efficient algorithms are developed for data classification tasks. These two novel algorithms combine the SVM and theGPC algorithms in a unified framework with a view to classifying large datasets ina relativelyshorter period of time.In addition, the proposed models provide probabilistic information on the classification which can be helpful to establish confidence intervals over the class outputs. All of these areachieved while ensuring that the classification accuracy remains within an acceptable level. Five model datasets have been used to perform the experimentation and examine the actual performance of the proposed algorithms in contrast with the existing state-of-the-art methodologies. The proposed algorithmshave proved to be computationally efficient, especially in the case of testing new data, where they consistently provided better results than the existing algorithms. The accuracies provided by the proposed algorithmsare also satisfactory, which at times, even surpass the accuracy level obtained by the existing algorithms.