A comparative study of various classification techniques in multivariate skew-normal data.

Abstract

The assumption of normality in data has been considered in the field of statistical analysis for a long time. However, in many practical situations, this assumption is clearly unrealistic. It has recently been suggested to study the performance of various statistical techniques like classification by using the data from distributions indexed by skewness/ shape parameters. In this study, four different classification techniques, namely linear discriminant analysis, quadratic discriminant analysis, k-th nearest neighbor and oblique axes method are considered for classification of observations. To assess the performance of the above techniques under non-normality caused by skewness, which is introduced in the ricebean data by using multivariate skew-normal distribution through simulation. Apparent error rate is used to study the classification performance of these techniques. The result of this study can be used for choosing the best method of classification for skewed-normal situation. The results indicate that k-th nearest neighbour followed by oblique axes method and linear discriminant analysis perform better in skew-normal situations than normal condition and quadratic discriminant analysis performed better in normal data. For maximum consistency and accuracy of classification of skew-normal data, k-th nearest neighbor is best among the four classification techniques