NkEL: nearest k-labelsets ensemble for multi-label learning

Multi-label learning (MLL) can be viewed as an extension of multi-class learning (MCL) that supports nonexclusive labels. Random k-labelset ensemble (RAkEL) is a popular algorithm that transforms MLL into a series of MCL tasks to exploit label correlations. However, its effectiveness is impacted by the randomness of labelset construction. In this paper, we propose an MLL algorithm with nearest k-labelsets ensemble (NkEL) possessing three techniques. First, we select a labelset with a size of k for each label using the nearest-neighbor technique. Thus, NkEL considers high-order label correlations and has strong adaptability. Second, for each MCL problem, we build a neural network to provide numerical rather than categorical predictions. Therefore, the output values represent the confidence levels of different classes. Third, we propose an intra-labelset ensemble strategy for each label. This approach alleviates the limitations imposed by low class separability with the support of the total probability theorem. Experiments are conducted on datasets derived from various domains to compare the proposed method with fourteen popular algorithms. The results obtained in terms of six ranking-based and two classification-based measures demonstrate the feasibility and effectiveness of NkEL. The source code is available at github.com/fansmale/nkel.