Augmented Chains to Ensemble of Classifier Chains

Abstract

Multi-label classification (MLC) problems, where instances are associated with multiple labels, are commonly employed in everyday applications. There are several approaches to solving MLC problems and the ensemble of classifier chains (ECC) is one such method used as the basis of this article. ECC uses a binary classifier for each label and creates a chain of these classifiers in a specific sequence. However, the method has issues related to the order of the chain and the number of labels. Many studies try to find the best chain order or reduce the number of labels to improve results. This article aims to evaluate whether the insertion of meta-labels, created from combinations of the original labels, can enhance ECC prediction results. The approach involves creating combinations of labels through similarity correlation, selecting the most relevant labels based on these correlations, incorporating them into the dataset, and subsequently evaluating the model and prediction results. Results obtained in experiments with 19 well-known multi-label datasets and evaluated with 12 different measures show that the proposed approach improves Micro-Precision, Precision, Hamming-Loss, and Subset-Accuracy.