Abstract
Patterns in a data set have different levels of usefulness to the training of classifiers. Extreme patterns which are on the surface of the space occupied by the data are essential to the successful classifier design in some applications such as novelty detection. This paper presents a method that selects extreme patterns by analysing local geometrical information. It specifies a tangent haperplane at each pattern in order to determine the location of the pattern in the pattern space. A pattern is selected as an extreme pattern if it is identified to sit on the tangent hyperplane. The proposed method is evaluated using multilayer perceptrons and support vector machines on three benchmark classification problems. The experimental results demonstrate the proposed method is able to select extreme patterns for successful classifier design.
Original language | English |
---|---|
Title of host publication | Unknown Host Publication |
Pages | 1259-1266 |
Number of pages | 8 |
Publication status | Published (in print/issue) - 11 Jun 2007 |
Event | The World Congress on Engineering Asset Management and The International Conference on Condition Monitoring - Harrogate, UK Duration: 11 Jun 2007 → … |
Conference
Conference | The World Congress on Engineering Asset Management and The International Conference on Condition Monitoring |
---|---|
Period | 11/06/07 → … |