IA-KNNR: A Novel Imbalance-Aware Approach for Handling Multi-Label Class Imbalance Problem

Multi-label learning (MLL) is a supervised learning where the classifier needs to learn from the data where one instance can belong to more than one class (label). Due to its wide application, MLL has attracted great attention from academia and researchers. However, due to the complex nature of multi-label data (MLD), it suffers from the class imbalance problem, where data distribution is heterogeneous, and learning from that imbalanced data is challenging. To address this, we propose IA-KNNR, a novel imbalance-aware resampling framework that leverages distance-weighted K-nearest neighbors, label similarity, and a dynamic weighting strategy that accounts for the co-occurrence patterns between frequent and infrequent labels. In addition, DBSCAN-based prototype selection is used to select the most appropriate training data to reduce the complexity of the training time. Extensive experiments on ten multi-label benchmark datasets demonstrate that IA-KNNR performs significantly better than state-of-the-art approaches, up to 26.4%.