L-VSM: Label-Driven View-Specific Fusion for Multiview Multilabel Classification.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Multiview multilabel (MVML) classification involves instances with heterogeneous features and multiple labels.
• Current MVML methods often rely on shared subspace representations, with limited effectiveness for label characterization.
• The efficacy of shared subspace representations in MVML remains an open research question.

Purpose of the Study:

• To propose a novel label-driven view-specific fusion method for MVML classification.
• To develop a method that directly encodes individual view features for classifier induction, bypassing shared subspace learning.
• To enhance the characterization of relevant labels in MVML tasks.

Main Methods:

• Proposed L-VSM (Label-driven View-Specific fusion) method for MVML classification.
• Constructed label-driven feature graphs for each view and integrated them into a unified graph.
• Employed graph attention mechanisms for feature node aggregation and update, encoding intra-view and inter-view information.
• Introduced transformer architecture for dynamic semantic-aware label graph construction.
• Utilized multilabel soft margin loss to derive instance-label affinity scores.

Main Results:

• L-VSM achieved superior performance compared to state-of-the-art methods in extensive MVML experiments.
• The proposed method effectively leverages view-specific features and label correlations.
• Experimental validation across various MVML applications confirmed the method's effectiveness.

Conclusions:

• The proposed L-VSM method offers a novel and effective approach to MVML classification.
• Bypassing shared subspace learning and directly encoding view-specific features enhances model performance.
• The integration of graph attention and transformers provides a robust framework for MVML tasks.