5:30pm – 6:15pm : Feature selection and data fusing with spikey curves.

Professor Phil Brown [University of Kent ]

Many modern instruments present continuous functions as data, sampled at regular adjacent points. There is a natural belief that such continuity can be exploited in statistical analysis, but this has proved surprisingly elusive. Most more pragmatic approaches to analysing such data rely on data reduction
techniques that all but destroy such continuity. For example in mass spectroscopy proteomics it is usual to reduce the intensity curves
to a series of peaks, which might correspond to particular important peptides. We review the traditional approach and explore an approach which retains the original curves as data and uses wavelets to exploit and combine different data streams, for example from different laser settings. By doing this we are able to reveal important peaks that would otherwise be masked by noise. I will illustrate with data from mass spectroscopy cancer studies.

6:15pm – 6:45pm : Break for Refreshments

6:45pm – 7:30pm : Chemometrics and Calibration in Near Infrared Spectroscopy

Professor Tom Fearn [Statistical Science UCL]

The standard calibration problem in quantitative near infrared spectroscopy involves deriving a rule for predicting some chemical or physical property of a sample, e.g. the protein or moisture content of a sample of wheat, from its near infrared spectrum. For this a training set of, say, n = 100 samples with known properties is used. Since the spectrum may be measured at p = 1000 wavelengths this is an example of a problem with p >> n, many more predictors than training cases. Such problems have become fashionable of late, because of technology such as the gene chip, but they have been studied in the context of spectroscopy for at least 25 years. I shall describe some of the solutions adopted, both old and new.

Registration required:

Yes

Free:

Yes