No abstract

The special emphasis of support vector machines (SVMs) on generalization ability makes this approach particularly interesting for real-world applications with limited amounts of training data. In this paper we analyse the applicational aspects of SVMs, illustrating them with the step-by-step construction of a classifier for polymers by means of their mid-infrared spectra. With this example we show how the main difficulties of a typical industrial classification task can be addressed using SVMs.

Outliers may hamper proper classical multivariate analysis, and lead to incorrect conclusions. To remedy the problem of outliers, robust methods are developed in statistics and chemometrics. Robust methods reduce or remove the effect of outlying data points and allow the 'good' data to primarily determine the result. This article reviews the most commonly used robust multivariate regression and exploratory methods that have appeared since 1996 in the field of chemometrics. Special emphasis is put on the robust versions of chemometric standard tools like PCA and PLS and the corresponding robust estimates of regression, location and scatter on which they are based.

Signatures of Bovine Spongiform Encephalopathy (BSE) have been identified in serum by means of ''Diagnostic Pattern Recognition (DPR)''. For DPR-analysis, mid-infrared spectroscopy of dried films of 641 serum samples was performed using disposable silicon sample carriers and a semi-automated DPR research system operating at room temperature. The combination of four mathematical classification approaches (principal component analysis plus linear discriminant analysis, robust linear discriminant analysis, artificial neural network, support vector machine) allowed for a reliable assignment of spectra to the class ''BSEpositive'' or ''BSE-negative''. An independent, blinded validation study was carried out on a second DPR research system at the Veterinary Laboratory Agency, Weybridge, UK. Out of 84 serum samples originating from terminally-ill, BSE-positive cattle, 78 were classified correctly. Similarly, 73 out of 76 BSE-negative samples were correctly identified by DPR such that, numerically, an accuracy of 94.4 % can be calculated. At a confidence level of 0.95 (a~0.05) these results correspond to a sensitivity w 85% and a specificity w 90%. Identical class assignment by all four classifiers occurred in 75% of the cases while ambiguous results were obtained in only 8 of the 160 cases. With an area under the ROC (receiver operating charateristics) curve of 0.991, DPR may potentially supply a valuable surrogate marker for BSE even in cases in which a deliberate bias towards improved sensitivity or specificity is desired. To the best of our knowledge, DPR is the first andup to now-only method which has demonstrated its capability of detecting BSE-related signatures in serum.

The problem of resolving bilinear two-way data into the contributions from the underlying mixture components is of great interest for all hyphenated analytical techniques. The fact that the optimal solution to this problem at least to some extent depends on the nature of the data under study has lead to a numerous different approaches. One of the seminal publications in this area was contributed by Olav M. Kvalheim and Yi-Zeng Liang in 1992. They not only provided valuable Heuristic Evolving Latent Projections (HELP) but also enlightened many important aspects of curve resolution in this and numerous subsequent publications. Here we extend their key concept of HELP, that is the use of latent projective graphs for identifying one-component regions, by using polar coordinates for these analyses and thereby creating a simple, intuitive exploratory tool for directly solving the curve resolution problem for two and three components graphically. Our approach is demonstrated with simulated data, an example from reaction monitoring with broadband ultrafast spectroscopy and one chemometric standard data set.