Support vector machine parameter optimization for text categorization problems

Abstract

This paper analyzes the influence of different parameters of Support Vector Machine (SVM) on text categorization performance. The research is carried out on different text collections and different subject headings (up to 1168 items). We show that parameter optimization can essentially increase text categorization performance. An estimation of range for searching optimal parameter is given. We describe an algorithm to find optimal parameters. We introduce the notion of stability of classification algorithm and analyze the stability of SVM, depending on number of documents in the example set. We suggest some practical recommendations for applying SVM to real-world text categorization problems.