This study investigates the combination of four different classification methods for text categorization through experimental comparisons. These methods include the Support Vector Machine, kNN (nearest neighbours), kNN model-based approach (kNNM), and Rocchio methods. We first review these learning methods and the method for combining the classifiers, and then present some experimental results on a benchmark data collection of 20-newsgroup with an emphasis of average group performance - looking at the effectiveness of combining multiple classifiers on each category. In an attempt to see why the combination of the best and the second best classifiers can achieve better performance, we propose an empirical measure called closeness as a basis of our experiments. Based on our empirical study, we verify the hypothesis that when a classifier has the high closeness to the best classifier, their combination can achieve the better performance.
|Title of host publication||Unknown Host Publication|
|Place of Publication||HEIDELBERGER PLATZ 3, D-14197 BERLIN, GERMANY|
|Number of pages||10|
|Publication status||Published - 2004|
|Event||DATABASE AND EXPERT SYSTEMS APPLICATIONS, PROCEEDINGS - Zaragoza, Spain|
Duration: 1 Jan 2004 → …
|Name||LECTURE NOTES IN COMPUTER SCIENCE|
|Conference||DATABASE AND EXPERT SYSTEMS APPLICATIONS, PROCEEDINGS|
|Period||1/01/04 → …|
Bi, YX., Bell, D., Wang, H., Guo, GD., & Dubitzky, W. (2004). Classification decision combination for text categorization: An experimental study. In Unknown Host Publication (pp. 222-231). (LECTURE NOTES IN COMPUTER SCIENCE)..