T. L. Isenhour

An algorithm for a selective nearest neighbor decision rule (Corresp.)

A procedure is introduced to approximate nearest neighbor (INN) decision boundaries. The algorithm produces a selective subset of the original data so that 1) the subset is consistent, 2) the distance between any sample and its nearest selective neighbor is less than the distance from the sample to any sample of the other class, and 3) the subset is the smallest possible.

The Evaluation of Mass Spectral Search Algorithms

Masa spectral search algorithms are tested by two methods. Trial searches of target spectra are used to compare the effects of various data-encoding methods and different distance metrics on search results. The effectiveness of selected search algorithms in retrieving similar compounds is assessed through the use of a propagation method, which employs repetitive searches of the nearest matches to a “seed” compound. A library containing the spectra of nearly 17 000 organic compounds is used with these methods to provide an evaluation of the relative performance of different search techniques.

Library Retrieval of Infrared Spectra Based on Detailed Intensity Information

Infrared spectral searches which rely on detailed intensity information and those based on peak position information are compared using a library of 500 reference spectra. Spectra with detailed intensity information are encoded with absorbance values for 200 resolution increments spanning the range from 4000 to 200 cm−1. Three different distance metrics are evaluated for searches using intensity information, and the effects of selecting data from limited regions of the infrared spectrum are examined.

Compression of an Infrared Spectral Library by Karhunen-Loeve Transformation

A data compression technique using factor analysis and the Karhunen-Loeve transformation has been applied to a library of vapor phase infrared spectra. Application of the compression algorithm results in a fivefold reduction in storage requirements for the spectral library and a corresponding reduction in search time. A search system based on the compression data has been evaluated and found to compare favorably with a search using the entire spectrum.

Infrared Library Search on Principal-Component-Analyzed Fourier-Transformed Absorption Spectra

Fourier-transformed infrared absorption spectra are compressed by principal-component analysis. Searches using the compressed data demonstrate the success of the compression. A search technique utilizing principal components in a prefilter selection produces the most efficient infrared library search method.

Investigation of the Fourier Transform for Analyzing Spectroscopic Data by Computerized Learning Machines

This paper investigates the use of the fast Fourier transform as an aid in the analysis and classification of spectroscopic data. The pattern obtained after transformation is viewed as a weighted average and/or as a frequency representation of the original spectroscopic data. In pattern recognition the Fourier transform allows a different (i.e., a frequency) representation of the data which may prove more amenable to linear separation according to various categories of the patterns. The averaging property means that the information in each dimension of the original pattern is distributed over all dimensions in the pattern resulting from the Fourier transformation. Hence the arbitrary omission or loss of data points in the Fourier spectrum has less effect on the original spectrum. This property is exploited for reducing the dimensionality of the Fourier data so as to minimize data storage requirements and the time required for development of pattern classifiers for categorization of the data. Examples of applications are drawn from low resolution mass spectrometry.

Principal component analysis of the infrared spectra of mixtures

Abstract Principal component analysis of the infrared spectra of a series of related mixtures is used to determine the number of compounds present. The use of empirical error estimates makes it possible to determine correctly the number of components even when the spectra of the individual compounds are very similar.

A Comparison of Two Discriminant Functions for Classifying Binary Infrared Data

Some form of information compression is essential if one is to be able to utilize effectively the increasingly large data compilations. One approach is to eliminate the intensity information, leaving spectra packed in a peak/no peak format. This paper reports the comparison of two simple discriminant functions for classifying binary infrared data. For the multicategory problem of 13 classes used in this investigation, random guessing would achieve about 8% correct classification. A dot product calculation produces 49.1% correct classification, while a distance measurement produces 58.7%. The results from this investigation are also qualitatively compared to previous work using infrared data which retained some intensity information. It is found that the binary packing of spectral data shows great promise in the area of infrared analysis.

An Infrared Search System Based on Direct Comparison of Interferograms

A compound identification system for gas chromatography/Fourier transform infrared spectroscopy data has been developed based on the direct comparison of interferograms. A multidimensional vector approach is used to form compound-specific vectors from the interferograms of mixture components. Comparisons are made with a set of known compound vectors by calculating how close together any two vectors lie in space. This system is used to identify the components of several synthetic mixtures.

Pattern Recognition Methods for the Classification of Binary Infrared Spectral Data

Five pattern recognition methods are compared for their ability to classify binary infrared spectra. Included is a discussion of the time vs success balance for each of the techniques. Predictive ability decreases in the order maximum likelihood > distance > Tanimoto similarity ∼ Hamming distance > dot product. The time required for each prediction after the classifier has been developed increases in order maximum likelihood ∼ distance ∼ dot product < Tanimoto similarity ∼ Hamming distance.