In the present work, the simplex optlmiratlon procedure is examined as a means of producing high quality linear discriminant functions for recognitlon of eleven functional group categories from low resolution mass spectra. Classifier performance Is evaluated with test sets totaling over 1900 compounds; and a new performance measure-the figure of merlt-Is employed In the evaluation. For linearly separable data, slmpiex-derived weight vectors are not superior to those obtained by the simpler error-correction-feedback methods. For inseparable data, however, a simplex with lnltial vertices at welght vectors derived by error-correction-feedback does indeed generally converge on a set of weight vectors with superior performance. It is found that those weight vectors showing poorest performance correspond to the more difficult chemical discriminations.Studies o f chemical pattern recognition have shown the promise of the method for eventual routine application to spectral analysis (1-4). Previous work has been open to the criticism that it was based on relatively small data sets unrepresentative of practical analytical problems. Therefore, a primary objective of the work presented here was to investigate the applicability of linear discriminant analysis to a larger data set. Accordingly, 2140 mass spectra drawn from a larger collection of 18 806 spectra ( 5 ) , were used in this study.
We have refrained from value judgments such as "good" or "better than" in discussing both the classifiers and the evaluators. The actual characteristics which make a classifier suitable for a particular application may depend on the application itself, as, for example, when the penalty associated with misclassification of a class member is not the same as for misclassification of a nonmember.24 Although the percent correct prediction should probably be abandoned as a measure of the performance of binary classifiers. the other measures discussed can be useful in developing a total picture of relative and absolute classifier performance.Abstract: Recent research on the use of adaptive networks of digital learning elements for chemical pattern recognition has stressed the high performance of such classifiers and their applicability to linearly inseparable data. In the present work, we apply a new performance measure, the figure of merit, and a large set of test data in a rigorous evaluation of the performance of digital learning networks. The results herein reported show that, when confronted with a large data set selected without particular consideration of the peculiarities of the network, the digital learning network continues to give good performance, although this performance is substantially below the levels previously reported. A comparison of the performance of the digital learning network classifiers with that of a set of linear discriminant functions indicates similar levels of performance for the two types of classifier.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.