Analyzing The Effect Of Feature Selection On K-Nearest Neighbor Performance In Hepatitis C Classification

Abstract

This study analyzes the effect of feature selection on the performance of the K-Nearest Neighbor (KNN) algorithm in classifying categories in the UCI Hepatitis C dataset, which contains 615 observations, 12 predictor features, and five target classes. Three scenarios were compared: KNN using all features, KNN with Chi-Square feature selection, and KNN with the RChiFE approach, operationalized through Tukey clipping, imputation, scaling, and Chi-Square feature selection. Evaluation used nested stratified cross-validation (outer 5-fold and inner 3-fold), with parameter selection based on macro-F1. The results show that RChiFE achieved a mean accuracy of 0.9187 (SD 0.0287) and macro-F1 of 0.5079 (SD 0.0973), compared with 0.9008 accuracy and 0.3882 macro-F1 for KNN using all features, and 0.8911 accuracy and 0.3818 macro-F1 for Chi-Square. These findings indicate improved average performance for RChiFE in this experiment, although recall for minority classes remains a concern.