Evaluation of SVM Transformations for Multi-label Research Article Classification

Support Vector Machine (SVM) models are often used to address the problem of multi-labeled topic classification across various datasets. However, to fit the SVM-required modeling, it is necessary to transform the multi-labeled dataset into single-labeled ones, which plays a crucial role in the performance. Different transformations applied to the problem may lead to varying accuracy and behavior of SVM over datasets. This paper investigates the performance of the SVM model on a multi-labeled research articles dataset, with three commonly used transformations: Binary Relevance (BR), Label Powerset, and Classifier Chains. It also performs a case study of categorizing research articles into one or more related topics. The results indicate that the Label Powerset transformation achieved the best average accuracy score across all topics classification. Moreover, the Label Powerset and BR transformations were able to achieve a Hamming loss measurement of ≈90% for the fraction of topics that are incorrectly assigned. However, BR exhibited the best recall and precision balancing for class classification measurements. The paper introduces the Least Class Classifier (LCC) technique, which challenges the problem of imbalanced datasets to achieve an equal chance for the minor classes. This technique addresses the problem of imbalanced datasets. It also shows promising results for increasing recall calculations for the minor class in imbalanced datasets. That emphasizes the performance of SVM models for multi-labeled topic classification can be improved by selecting an appropriate transformation technique.