One‐class classifiers

Abstract

The principles of one-class classifiers are introduced, together with the distinctions between one-class/multiclass, soft/hard, conjoint/disjoint and modelling/discriminatory methods. The methods are illustrated using case studies, namely from nuclear magnetic resonance metabolomic profiling, thermal analysis of polymers and simulations. Two main groups of classifier are described, namely statistically based distance metrics from centroids (Euclidean distance and quadratic discriminant analysis) and support vector domain description (SVDD). The statistical basis of the D statistic and its relationship with the F statistic, χ 2 , normal distribution and T 2 is discussed. The SVDD D value is described. Methods for assessing the distance of residuals to disjoint principal component models (Q statistic) and their combination with distance-based methods to give the G statistic are outlined.

Pattern recognition in chemometrics

“…A few chapters do mention techniques that we would consider relevant to PR in a few other chapters, but the profile of PR is low. In addition it is rarely acknowledged that MSPC (multivariate statistical process control) is in practice a form of PR via one class classifiers [36], and few people felt that to safely use most classification techniques one has to understand the basics of distributions: SIMCA for example is not usually introduced this way.…”

“…Yet most users understand little about its basics. Few appreciate that it is in fact a one class classifier [36,50] and so is based on a radically different philosophy to, for example, LDA or PLS-DA. It cannot be compared directly to two class or multiclass classifiers.…”

Review of validation and reporting of non-targeted fingerprinting approaches for food authentication

“…Classification approaches might be conceptually separated in those that model each class separately, often referred to as oneclass classifier or class modelling, and those that attempt to discriminate each class using one model in a so-called discriminant analysis [71,72]. PLS-DA is an example for a discriminant analysis approach often used in the investigated studies.…”

Comparison of classification methods with “n-class” receiver operating characteristic curves: A case study of energy drinks

“…SIMCA belongs to the group of class modeling techniques; some authors term it one-class classification. [22] "The application of an algorithm to different types of data may result in diverse outputs, and in general, no single algorithm is optimal for solving all problems. Each set of chemical data therefore requires the choice of an optimal (or a near-optimal) algorithm for that particular data set" [23].…”