Analisis Pemeringkatan Kualitas Klasifier Pada Dataset Tidak Seimbang

Abstract

Classification algorithms C4.5, CART, k-Nearest Neighbors (k-NN) and Naive Bayes are included in the "Top 10 algorithms in data mining". The author tests and analyzes the four to get a ranking order according to the quality of performance. A common method to compare the quality of classifier performance for classifying two class labels with a balanced the proportion of the number of classes of datasets is to test the performance of classifier accuracy. For unbalanced datasets such as in this study using this method can be biased, and can even lead to misleading conclusions. By calculating the scores that are a combination of performance parameters "accuracy", "precision", "recall" and "AUC" where the highest value of each parameter is the best will result in a more representative rating of the classifier's performance indicating the quality of the classifier. Two test methods are used, namely 10-fold Cross Validation and Discrete Testing to ensure the results of a representative performance evaluation of each classifier. The implementation of the testing of the four classification algorithms above and the comparative analysis of the performance results in the ranking of the best quality performance ratings, namely: 1. k-NN, 2. C4.5, 3. CART, 4. Naive Bayes.