Analisis Pengaruh Seleksi Atribut Relief-F dan Gain Ratio Terhadap Performa Naïve Bayes Classifier

Abstract

Naïve Bayes Classifier (NBC) is one of the most popular probabilistic classification algorithms in 
data mining, known for its simplicity and efficiency. However, NBC performance tends to 
degrade when datasets contain irrelevant or noisy attributes. This study analyzes the effect of 
attribute selection using Relief-F and Gain Ratio methods on the performance improvement of 
NBC. Two benchmark datasets from the UCI Machine Learning Repository were selected to 
represent contrasting data characteristics: the House Vote dataset (435 records, symbolic 
attributes with balanced class distribution) and the Bank Marketing dataset (45,211 records, 
numeric and categorical attributes with severe class imbalance, approximately 88% majority 
class). All experiments were implemented in Google Colaboratory using Python. Three 
experimental scenarios were applied to each dataset: (1) NBC without attribute selection as 
baseline, (2) NBC with Relief-F attribute selection, and (3) NBC with Gain Ratio attribute 
selection. Performance evaluation used 10-fold cross-validation with metrics including 
accuracy, precision, recall, F1-score, and confusion matrix. Results show that on the House Vote 
dataset, Relief-F increased NBC accuracy from 90.11% to 93.79% (+3.68%), while Gain Ratio 
reduced accuracy to 89.43%. On the Bank Marketing dataset, Relief-F improved accuracy to 
89.36% and improved minority class recall from 29.34% to 35.71%, while Gain Ratio yielded 
only marginal improvement. Overall, Relief-F proved more effective than Gain Ratio in enhancing 
NBC performance, particularly on datasets with clear classification patterns and imbalanced 
class distribution.