Additional Funding Sources
Work supported by the National Science Foundation under grant No. CHE-1506417 (co-funded by CDS and E Programs) and is gratefully acknowledged by the authors.
Abstract
A useful application in analytical chemistry is classifying unknown samples into classes. Single-class classification is a type of classification approach where only one well-defined class is of interest. Outlier detection is useful for defining class membership for unknown samples, since outlier detection removes samples that are not represented by the sample class space. When using outlier detection, there are two problems: which outlier measure to use and the tuning parameter value for the chosen outlier measure. The proposed technique for single-class classification using outlier measures eliminates these two problems. To avoid selecting any one particular outlier measure, multiple measures are evaluated by using sum of ranking differences (SRD). The method of SRD is used to evaluate multiple outlier measures to obtain a consensus in classifying a sample. In regards to tuning parameters, a parameter window is used to avoid doing more work, such as having a training set of samples to select a tuning parameter. Wavelength selection and fusing spectra from different instrument is used in conjunction with SRD to provide a robust characterization of the class of interest. Presented are results for the new classification approach on spectral food data sets.
Classification using Sum of Ranking Differences of Outlier Measures
A useful application in analytical chemistry is classifying unknown samples into classes. Single-class classification is a type of classification approach where only one well-defined class is of interest. Outlier detection is useful for defining class membership for unknown samples, since outlier detection removes samples that are not represented by the sample class space. When using outlier detection, there are two problems: which outlier measure to use and the tuning parameter value for the chosen outlier measure. The proposed technique for single-class classification using outlier measures eliminates these two problems. To avoid selecting any one particular outlier measure, multiple measures are evaluated by using sum of ranking differences (SRD). The method of SRD is used to evaluate multiple outlier measures to obtain a consensus in classifying a sample. In regards to tuning parameters, a parameter window is used to avoid doing more work, such as having a training set of samples to select a tuning parameter. Wavelength selection and fusing spectra from different instrument is used in conjunction with SRD to provide a robust characterization of the class of interest. Presented are results for the new classification approach on spectral food data sets.