|
Abstract:
|
Microarray (and other high throughput omics) data are characterized as having an incredibly large number of predictor variables, only a few of which are truly informative and are correlated in groups, and a quite small number of samples. Classification methods, when used in conjunction with a conventional dissimilarity measure, tend to have terribly low efficiencies when applied to such data. We propose a novel dissimilarity measure that overcomes this problem. The measure is obtained via an ensemble scheme that involves a series of resample classifications, where the resampling involves both predictors and samples and is weighted towards informative features of the predictor space. We demonstrate, using actual microarray data, that this produces far superior classifications and visualizations of the data.
|
