ECC-CS: a multi-label data classification algorithm for class imbalance based on cost-sensitive learning

Abstract

Multi-label learning has attracted much attention due to its numerous applications in the real world. Although many multilabel learning algorithms have been proposed, most of them neglect the class imbalance problem that exists in multi-label data. Even when some studies focus on this problem, they often fail to explore label correlations. To address these two issues simultaneously, a cost-sensitive-based class imbalance multi-label data classification method called ECC-CS (Ensemble classifier chains - Cost sensitive) is proposed. In brief, ECC-CS combines the popular Ensemble Classifier Chain (ECC) algorithm with cost-sensitive learning techniques. Therefore, the ECC-CS algorithm simultaneously inherits the merit of label correlation exploration held by ECC and the merit of addressing the class imbalance problem owned by cost-sensitive learning techniques. Specifically, in ECC-CS, the cost-sensitive learning technique runs on each branch of ECC. 12 benchmark multi-label datasets were used to compare the proposed ECC-CS algorithm with several traditional class imbalance multi-label learning algorithms. The experimental results have indicated the effectiveness and superiority of the proposed ECC-CS algorithm.