SVM Paradoxes

• Published in: Perspectives of Systems Informatics pp. 86 - 97
• Main Authors: Beney, Jean, Koster, Cornelis H. A.
• Format: Book Chapter
• Language: English
• Published: Berlin, Heidelberg Springer Berlin Heidelberg 2010
• Series: Lecture Notes in Computer Science
• ISBN: 3642114857; 9783642114854
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-642-11486-1_8

Support Vector Machines (SVM) is widely considered to be the best algorithm for text classification because it is based on a well-founded theory (SRM): in the separable case it provides the best result possible for a given set of separation functions, and therefore it does not require tuning. In this paper we scrutinize these suppositions, and encounter some paradoxes. In a large-scale experiment it is shown that even in the separable case SVM’s extension to non-separable data may give a better result by minimizing the confidence interval of the risk. However, the use of this extension necessitates the tuning of the complexity constant. Furthermore, the use of SVM for optimizing precision and recall through the F function necessitates the tuning of the threshold found by SVM. But the tuned classifier does not generalize well. Furthermore, a more precise definition is given to the notion of training errors.