Study K-NN Based Classifier To Overcome The Negative Effect Of Outliers

Abstract

This thesis studies the k-NN based classifier to overcome the negative effect of the outliers. Sometimes data sets always exist values that are out of range and different from the rest, these values are often called outliers. Outliers affect the training of the model and give incorrect prediction results. This thesis aims to research and analyze the methods to deal with this trouble and k-NN classifier was used to overcome the outliers in the data sets. Methods developed based on the k-NN classifier including k-GNN, LMkNN and LMkGNN were used to obtain the best results. Through experiments on five data sets from UCI Repository, the results illustrated that the LMkGNN classifier gave the highest accuracy of the four classifiers, and the ability to overcome the negative effect of outliers of the LMkGNN was also higher than the three sets remaining