An efficient self-paced multi-view method for partial label learning

doi:10.1016/j.knosys.2026.116969

An efficient self-paced multi-view method for partial label learning

2026-09-06

0

PRE

AI

DOI：10.1016/j.knosys.2026.116969

Original

Original request for help

Save

Abstract

En 中文

Partial label learning (PLL) aims to learn from ambiguously supervised data, where each training instance is associated with a candidate label set containing the ground-truth label. Although multi-view data can provide complementary information for resolving label ambiguity, existing multi-view PLL methods still struggle to jointly address label ambiguity, view heterogeneity, and progressive disambiguation within a unified framework. To this end, we propose SP-MVPLL, an efficient self-paced multi-view method for partial label learning. Specifically, we formulate multi-view PLL under a max-margin framework with latent labels and introduce a consensus distribution to enforce cross-view consistency at the output level. To alleviate the adverse effects of label ambiguity and noisy supervision, self-paced learning is incorporated to gradually emphasize easy and reliable instances during training. Meanwhile, an entropy-regularized adaptive view-weighting mechanism is developed to automatically balance the contributions of different views and prevent degenerate view aggregation. Extensive experiments on five benchmark datasets under different ambiguity levels demonstrate the effectiveness of SP-MVPLL against state-of-the-art baseline methods. A controlled two-view runtime study further shows that SP-MVPLL is the fastest method on three of the five datasets. When all available views are used, SP-MVPLL is faster than CFDM on every dataset, reducing training time by 52.4%–66.6%.

Keywords:

Partial label learning

Multi-view learning

Self-paced learning

Adaptive view weighting

An efficient self-paced multi-view method for partial label learning

Abstract

Journal

Organization

Cited Papers