Instance-Based Classification of Noisy Infrared Spectra.
Abstract
Successful systems for classification of real-world data must be tolerant of noise-that is, distortions introduced into the system's model of the real-world domain. Most classification systems are trained on a set of exemplars to identify features of each category and then tested on previously unseen instances. In an instance-based classification system using k-nearest neighbor (k-NN), the training phase is reduced to storing one or more exemplars for each category. During testing, a distance metric is applied to the features of the new instance to determine the k closest exemplars. A voting scheme assigns the category of the modal average to the testing instance. Unlike other methods, k-NN does not try to distinguish between 'relevant' and 'irrelevant' features. Nonetheless, k-NN has been shown to asymptotically approach optimal Bayesian accuracy. This report presents the results of applying k-NN to the problem of classifying chemical agents from noisy infrared absorption spectra (from a suite of chemical agents used elsewhere in the literature). Straightforward nearest-neighbor approaches without editing appear to be tolerant of random noise when the amounts of noise in the training and testing sets are relatively close. Performance of k-NN versus 1-NN approaches can be improved if the training sets are edited so as to exclude degenerate outliers and redundant positive instances.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jan 01, 1997
- Accession Number
- ADA321850
Entities
People
- Robert P . Winkler
- Timothy C. Gregory
Organizations
- United States Army Research Laboratory