Ranking Machine Learning Classifiers Using Multicriteria Approach

Classification algorithms are widely used as data mining tools for knowledge extraction. The literature presents several classifiers, but none of them applies to all problems. encountered in the various context in which they are used. Faced with this situation, the present article proposes a multicriteria approach to help practitioners to select the classifiers that will generate the best quality results by observing their performance measures. An empirical study was performed using a baseline of fetal examination from an UCI database using five classification algorithms (C4.5, Naive Bayes, SMO, KNN and Bayesnet), and each classifier was measured using five performance indicators (accuracy, true positive rate, precision, ROC curve and f-measure). Once implemented, a classifier ranking was conducted based on MCDA PROMETHEE II method, and the results show that SMO, C4.5 and Naive Bayes achieved the highest overall ranking.