Support vector machines: A Critical empirical evaluation

Abstract

The main focus of this thesis is Support Vector Xlachines (8\-:'1s). In particular, we first investigate empirically whether they live up to the claims made, i.e. they suffer less from class imbalances, noise and rnisclassifications in the data; they suffer less from overfitting and local minima and that 8V~\'1sare easier to tune than other machine learning algorithms. Secondly,__"I'.~e_have compared SVMs with other machine -learning ~igorithms on some benchmark datasets, Thirdly, we investigate whether S\ -?lIs can be used to filter an original dataset which is subsequently used to train a second machine learning algorithm, The hypothesis is that a reduced dataset made up of support vectors contains sufficient examples to tune in an optimal way. a second machine learning algorithm. Our findings here show that this is possible. However. the machine learning algorithm that is used to learn from this reduced dataset should be carefully selected. Fourthly, we have set up a methodology to tune and to compare different learning algorithms.