Multi Label Classification Based on Ensemble Learning: Investigative Study

In the big data technology nowadays, activities associated with Multi-Label Classification (MLC) pose big and complex challenges, receiving terrific interest in diverse fields. Existing MLC algorithms suffer from low overall performance. This paper considers four well-known Problem Transformation Methods (PTMs). Two of these PTMs are based on the frequency of the labels (MFL and LFL) while the last two PTMs are based on the positive correlations among labels. These PTMs have been implemented and compared against four multi label datasets of various sizes to identify the best one that fits the nature of MLC datasets. Moreover, six popular ensemble learning models have been carried out and compared on the four datasets with respect to several evaluation metrics such as Accuracy, Hamming loss, and Exact Match and using SMO, DecisionStump, Random Forest, and Random Commission classifiers. The results revealed that MFL accomplished the quality performance amongst the four PTMs, and Stacking SVM is the best ensemble learning technique.