Label recovery and label correlation co-learning for multi-view multi-label classification with incomplete labels

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Multi-view multi-label learning (MVML) faces performance degradation due to incomplete labels and unaddressed relationships between views and labels.
• Existing MVML algorithms often overlook the intricate correlations among labels and the interactions between different data views.

Purpose of the Study:

• To propose a novel method, MV2ML, for multi-view multi-label classification that effectively handles incomplete labels.
• To simultaneously recover incomplete label matrices and learn label correlations to improve classification accuracy.

Main Methods:

• Constructed label correlation-guided binary classifiers for each label.
• Employed multi-kernel fusion to integrate multi-view data, leveraging individual and complementary information.
• Developed a collaborative learning strategy for simultaneous label correlation exploitation, multi-view data fusion, label recovery, and classification model construction.

Main Results:

• MV2ML demonstrated highly competitive classification performance against state-of-the-art approaches.
• The proposed method achieved superior results on various real-world multi-view multi-label datasets.
• Experimental results were evaluated using six standard performance metrics.

Conclusions:

• The proposed collaborative learning strategy effectively addresses label incompleteness in MVML.
• MV2ML's simultaneous recovery of label matrices and learning of label correlations interactively boost classifier training.
• The method shows significant potential for improving multi-view multi-label classification tasks.