Document classification with support vector machine

In this thesis, document classification task is studied using an online Support Vector Machine. Lasvm algorithm is adapted as to work with pairwise kernels. At first, the effects of the choice of the kernel function and its parameters are considered. In order to classify documents with a linear decision bound, the data is mapped into a higher dimensional space. Therefore the data mapping procedure and the kernel trick are explained in detail with several examples. After the determination of optimal hyperparameters, experiments are conducted on different combinations of pairwise training and testing data. Performances of the SVM models are compared according to the classification accuracy and area under the ROC curve. Results indicate that online pairwise classification is a good alternative to the methods used in binary and multiclass classification tasks. In this document classification task, linear pairwise kernels achieve better results than the gaussian pairwise kernels because of the high dimensionality of the data at hand.