EFFECTS OF DATA DELETION AND WEIGHTING ON FISHER’S LINEAR CLASSIFICATION METHOD: A ROBUSTIFICATION APPROACH

Abstract

The classical supervised classification model's performance is hampered by the effects of influential observations (IOs). The influential observations(IO’s) when deleted, weighted,Winsorized, truncated and retained have enormous effects in making replicative inferences in different classification models. Due to the influence of IO’s on classical supervised classification models, different methods such as IO deletion or weighting have been introduced to reduce the influence of IO’s. Some of these influential observations reduction or deletion methods have resulted in information loss of various degrees. In this study, we investigated the effects of IOs deletion and weighting using the Mahalanobis distance as a plug in to enhance the robustness of the Fisher linear classification method (FLCM). We proposed an F-weight plug in method to robustify the FLCM. We compared the performance of these methods to determine whether IO deletion or IO weighting retards or enhances the classification accuracy of the FLCM. The study affirmed that IO weighting using the F-weight minimizes information loss more than the IO deletion using the Mahalanobis distance. This study concludes that the variant of FLCM based on the F-weight method showed improved classification accuracy, and efficiency more than the Mahalanobis distance based FLCM.