Research map / Label-efficient and robust learning / Noisy labels

Limited-Supervised Multi-Label Learning with Dependency Noise

AAAI 2024 · Label-efficient and robust learning › Noisy labels · approach: Instance-dependent noise modeling · 5 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes MLDN, a linear multi-label learner that models instance- and label-dependent label noise (missing and redundant labels) via group-sparse factorized noise with a manifold regularizer, solved by proximal first-order optimization.

Problem

Limited-supervised multi-label learning, where training labels contain missing and/or redundant labels whose noise is assumed independent of features and classes, which is unrealistic in practice.

Approach

Decompose the observed label matrix into true labels plus a noise matrix that is a linear function of features and labels (E = XP + YQ), with l2,1 group sparsity on P and Q to capture instance- and label-dependent noise, plus a graph-Laplacian manifold constraint on the noise; optimized with accelerated proximal gradient and closed-form updates.

Key results

• On partial multi-label data, MLDN ranks first in 83.3% of 24 dataset-metric comparisons on synthetic datasets and is best on real datasets m_emotion and mirflickr.
• On missing-label data, MLDN significantly outperforms MLMLFS, D2ML-L, D2ML-NL, MAXIDE and GLOCAL on all metrics and datasets.
• MLDN runs in 107.5 s on eurlex_dc, versus 42945 s for fPML in the PML setting.

Measured effects reported in the paper

Surprising findings

• The theory suggests the regularization parameters beta and gamma do not depend on the noise rate, so they need not be tuned to it.

Limitations

• Model is a linear regression predictor, so nonlinear feature learning is not explored.
• Experiments rely on synthetic noise injected into standard datasets with a fixed 20% missing and 20% redundant rate.
• Convergence is stated only as at least sublinear, and no hardware details are given.

Experimental setup

Datasets: slashdot medical enron scene yeast 20ng corel5k mirflickr eurlex_dc m_emotion

Benchmarks: slashdot medical enron scene yeast 20ng corel5k mirflickr eurlex_dc m_emotion

Metrics: Ranking Loss One Error Average Precision running time

Baselines: fPML PML-NI PART-VLS PART-MAP PML-LC PML-FP LMNNE D2ML-L D2ML-NL LEML MLMLFS GLOCAL

Builds on: partial multi-label learning multi-label learning with missing labels group lasso accelerated proximal gradient

Closest papers by meaning

• Noise-Robust Learning from Multiple Unsupervised Sources of Inferred Labels (AAAI 2022 · similarity 0.88)
• Partial Multi-Label Learning with Noisy Label Identification (AAAI 2020 · similarity 0.88)
• Learning with Feature-Dependent Label Noise: A Progressive Approach (ICLR 2021 · similarity 0.88)
• Partial Multi-Label Learning by Low-Rank and Sparse Decomposition (AAAI 2019 · similarity 0.88)
• Revisiting Sparsity Constraint Under High-Rank Property in Partial Multi-Label Learning (CVPR 2026 · similarity 0.87)
• Label Noise: Ignorance Is Bliss (NeurIPS 2024 · similarity 0.86)