Landmark-based Partial Multi-label Learning with Noise Processing
Boyuan Zhang, Zheming Li, Landong Liu, Zhenwu Wang · 2023
Multi-label learning (MLL) assumes that all labels are ground-truth, which can be costly or difficult to implement in practice. Partial multi-label learning (PML) provides an alternative by recognizing that each instance corresponds to a set of candidate labels, with only one subset representing the ground-truth label set. However, accurately identifying the ground-truth labels from the candidate set, which may be contaminated with noise, is the main challenge of PML. To address this challenge, we propose a landmark-based PML approach called LbPML, which incorporates noise recognition and space structure processing. We evaluate LbPML against three PML algorithms and three MLL algorithms using datasets from six different domains, and assess its performance using five commonly used evaluation criteria. Our extensive experimental results provide compelling evidence of the effectiveness of our proposed method.