Granular Ball-Guided Disambiguation for Partial Multilabel Feature Selection via Maximum Consistency Minimum...

Area of Science:

• Machine Learning
• Data Mining
• Artificial Intelligence

Background:

• Partial multilabel feature selection (PMLFS) addresses noisy labels in multilabel learning (MLL).
• Existing methods struggle with false positive labels and often rely on error-prone topology information or linear models.
• There's a need for methods that explore local data structures and effectively disambiguate labels.

Purpose of the Study:

• To propose a novel two-stage PMLFS method using granular computing.
• To enhance the accuracy of feature selection in the presence of noisy and partial labels.
• To overcome limitations of existing PMLFS approaches by incorporating label-specific information and local structure analysis.

Main Methods:

• A two-stage approach combining label disambiguation and feature selection.
• Stage 1: Label disambiguation using a granular ball computing model to capture label-specific information from data distribution.
• Stage 2: Filter-based feature selection utilizing a fuzzy decision neighborhood rough set (FDNRS) to explore local sample structures and minimize label uncertainty.

Main Results:

• The proposed method effectively disambiguates partial labels by analyzing data distributions.
• The FDNRS approach enhances feature selection by considering local structures and label relationships.
• Experiments on 12 datasets demonstrated the superior effectiveness of the proposed PMLFS approach across four evaluation metrics.

Conclusions:

• The novel two-stage PMLFS method significantly improves performance in multilabel learning with noisy labels.
• Granular computing and fuzzy rough sets offer robust solutions for label disambiguation and feature selection.
• The approach provides a more effective way to handle imprecise annotations and enhance MLL model accuracy.