Regularized Matrix Factorization for Multilabel Learning With Missing Labels.

Area of Science:

• Machine Learning
• Data Mining
• Computer Science

Background:

• Multilabel learning involves assigning multiple labels to data instances.
• Missing labels present a significant challenge in multilabel learning.
• Existing methods often rely on low-rank assumptions for label correlation recovery.

Purpose of the Study:

• To propose a novel approach for multilabel learning with missing labels.
• To improve the recovery of the ground-truth label matrix.
• To jointly optimize label recovery and multilabel classification model construction.

Main Methods:

• Regularized matrix factorization is employed to recover the ground-truth label matrix.
• Instance latent factors are regularized using local topological structure from feature space.
• Label latent factors and correlations are mutually adapted via label manifold regularization.

Main Results:

• The proposed method effectively recovers the ground-truth label matrix.
• Instance topological structure aids in inducing an effective multilabel model.
• Joint optimization benefits both label recovery and model construction.

Conclusions:

• The proposed regularized matrix factorization approach significantly outperforms state-of-the-art algorithms.
• The method demonstrates strong performance on both full-label and missing-label datasets.
• This work offers an effective solution for multilabel learning with missing labels.