The Effect on Classification Error of Random Permutations of the Features in Representing Multivariate Data by Faces.

摘要

A graphical method of representing multivariate data consists of having a computer draw a cartoon of a face which is determined by 18parameters (features) including length of nose,curvature of mouth,slant of eyes,length of eyes,etc. If a sample of 8-dimensional vector observations is presented each component of a vector can be made to determine one of the 18features and 10constants can be selected for the remaining features. The resulting output is a series of faces,one for each 8dimensional observation,which can be studied visually. An experiment was designed to evaluate the effect of a random permutation of the features on the visual ability to classify observations from two multivariate populations into two separate groups corresponding to the original populations. It is estimated that a random permutation may affect the error rate in this classification task by about 25%. (Author)