Partial Multi-Label Learning via Credible Label Elicitation.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Science

Background:

• Partial multi-label learning (PML) addresses scenarios with incomplete supervision where training examples have candidate labels, but only a subset is valid.
• A key challenge in PML is mitigating the impact of false positive labels within the candidate set, which can mislead model training.
• Existing methods struggle to effectively filter these noisy labels, hindering accurate multi-label prediction.

Purpose of the Study:

• To propose a novel two-stage approach for partial multi-label learning (PML) that enhances model induction by eliciting credible labels.
• To overcome the difficulty posed by false positive labels in PML by developing a robust label selection mechanism.
• To improve the generalization performance of multi-label predictors in the presence of inaccurate supervision.

Main Methods:

• A two-stage approach is proposed: 1) Estimating the labeling confidence of candidate labels using iterative label propagation.
• 2) Inducing a multi-label predictor by leveraging high-confidence credible labels via pairwise label ranking and virtual label splitting or maximum a posteriori (MAP) reasoning.

Main Results:

• The proposed approach demonstrates highly competitive generalization performance on PML tasks.
• Credible label elicitation effectively excludes a significant portion of false positive labels from the training process.
• The method successfully improves the accuracy of multi-label prediction by focusing on reliable labels.

Conclusions:

• The novel two-stage PML method effectively addresses the challenge of false positive labels through credible label elicitation.
• This approach significantly enhances the performance and robustness of multi-label learning under partial supervision.
• The findings suggest a promising direction for developing more accurate and reliable machine learning models in real-world scenarios with noisy data.