Comparative Linear Classification Splicing

Fred Zin

doi:10.24297/jam.v13i4.6258

Abstract

The conventional Fisher linear classification analysis has been investigated by numerous researchers and this has led to different modification or splicing due to non- robustness when the assumptions are violated and also when the data set contains influential observations. This paper adduced a winsorized procedure to robustify the probability base classification approach. The comparative classification performance of the Fisher linear classification analysis and its spliced versions when the data set are contaminated are investigated. The simulation results revealed that the robust Fisher's approach based on the minimum covariance determinant estimates outperformed the other procedures; a good competitor to this technique is the winsorized probability base classification technique. Though, the robust Fisher's technique using the minimum covariance determinant estimates breakdown for mixture contamination. On a general note, the conventional Fisher's approach and the probability base technique performed comparable.

Highlights

Linear classification techniques such as the Fisher’s method have been studied extensively and its modifications have no restriction in multivariate statistics
The Fisher’s method is basically designed in such a way that maximizing the between group scatter and minimizing the within group scatter allows for maximum separation of the groups
The conventional probability base classification rule discussed previously, utilized the conventional sample mean which may be susceptible to influential observations

Summary

Introduction

Linear classification techniques such as the Fisher’s method have been studied extensively and its modifications have no restriction in multivariate statistics. The robustified sample mean vectors and covariance matrices are plug-in into the conventional multivariate procedures to obtain robust multivariate techniques including the Fisher’s technique. The under classification performance of the Fisher linear classification rule is due to estimation errors of the mean vectors and covariance matrices(Pohar 2004).

Results

Conclusion