Learning embedded label-specific features for partial multi-label learning

Abstract

Partial multi-label learning (PML) aims to learn from instances with weak supervision, where each instance is associated with a set of candidate labels, among which only a subset is valid. Most existing approaches rely on identical feature representations to distinguish all class labels, overlooking the inherent distinctiveness of different labels, which leads to suboptimal model performance. Although recent studies have attempted to address this limitation by tailoring label-specific features, critical shortcomings remain: (1) isolated processing of feature tailoring and label disambiguation fails to leverage their synergistic relationship, and (2) direct extraction of label-specific features from the original feature space tends to yield unreliable results due to inherent noise and disturbances. This paper proposes a unified PML framework that jointly performs label disambiguation, embedded label-specific feature learning, and model induction. Within this framework, identifying ground-truth labels and generating label-specific features mutually reinforce each other, leading to continuous refinement. By customizing features from a compact and noise-free embedded space, the framework further ensures robustness and reliability in learning. Specifically, low-rank and sparse decomposition is employed to separate ground-truth labels from noisy ones, while a linear embedding discriminant model simultaneously generates embedded label-specific features and induces the model. Moreover, we enhance the classifier’s accuracy by assuming that the input and output spaces share local geometric structures, encouraging similar instances to have similar label sets. Extensive experiments on sixty-six real-world and synthetic datasets demonstrate that the proposed approach significantly outperforms state-of-the-art methods.