On Feature Selection with Measurement Cost and Grouped Features
Feature selection is an important tool reducing necessary feature acquisition time in some applications. Standard methods, proposed in the literature, do not cope with the measurement cost issue. Including the measurement cost into the feature selection process is difficult when features are grouped together due to the implementation. If one feature from a group is requested, all others are available for zero additional measurement cost. In the paper, we investigate two approaches how to use the measurement cost and feature grouping in the selection process. We show, that employing grouping improves the performance significantly for low measurement costs. We discuss an application where limiting the computation time is a very important topic: the segmentation of backscatter images in product analysis.
KeywordsFeature Selection Discrete Cosine Transform Feature Subset Feature Selection Algorithm Measurement Cost
- 1.C.L. Blake and C. J. Merz. UCI repository of machine learning databases, 1998. http://www.ics.uci.edu/mlearn/MLRepository.html.
- 2.F. Ferri, P. Pudil, M. Hatef, and J. Kittler. Comparative study of techniques for large-scale feature selection, 1994.Google Scholar
- 3.Anil Jain and Douglas Zongker. Feature selection: Evaluation, application, and small sample performance. IEEE Trans. Pattern Analysis and Machine Inteligence, 19(2):153–158, February 1997.Google Scholar
- 4.Pavel Paclík, Robert P.W. Duin, and Geert M.P. van Kempen. Multi-spectral Image Segmentation Algorithm Combining Spatial and Spectral Information. In Proceedings of SCIA 2001 conference, pages 230–235, 2001.Google Scholar
- 5.Pavel Paclík, Robert P.W. Duin, Geert M.P. van Kempen, and Reinhard Kohlus. Supervised segmentation of backscatter images for product analysis. accepted for International Conference on Pattern Recognition, ICPR2002, Quebec City, Canada, August 11–15, 2002.Google Scholar