On 10 June 2014 the RSS Merseyside local group, together with the RSS General Applications Section (GAS), hosted an afternoon of talks relating to Supervised Classification. Sixty-two people attended the meeting, held at the University of Liverpool, of which a quarter were RSS members. The session attracted visitors from Glasgow Caledonian University, the Health and Safety Executive, Plymouth University, University of Sheffield, and AstraZeneca, in addition to the University of Liverpool.
Dr Gabriela Czanner from the University of Liverpool began the afternoon with a brief introduction into supervised classification methods. Having defined classification and described examples where it is used, Gabriela went on to discuss details of Fisher’s method of classical linear discriminant analysis before giving an overview of other classification methods and of current trends in their development.
Professor David Hand, a senior research investigator at Imperial College London, then proceeded to discuss the quality of classification rules. Much to the consternation of the audience, David examined properties of common measures of classification performance and came to the conclusion that the area under the curve (AUC) should not always be used for evaluating classification rules! He did agree, however, that in some situations AUC is good to use; for example in cases where receiver operating characteristic (ROC) curves do not cross each other. David concluded that although some measures are favoured in the literature, there is little evidence that the measures are chosen to reflect the particular objectives of classification in the domain in question.
After a well-deserved coffee break Dr Marta Garcia-Finana from the University of Liverpool discussed whether principal component analysis is a good idea or not to reduce the dimensionality of a dataset, using a number of biomedical examples. To conclude the afternoon, Dr Christian Hennig from University College London introduced the audience to quantile classifiers. They are defined by classifying an observation according to a sum of appropriately weighted component-wise distances of the components of the observation to the within-class quantiles. He concluded that a quantile-based classifier with skewness adjustment is a promising classifier for high-dimensional data.