End-to-End Feature-Aware Label Space Encoding for Multilabel Classification With Many Classes.

Area of Science:

• Machine Learning
• Computer Science
• Artificial Intelligence

Background:

• Multilabel classification with numerous classes presents computational challenges.
• Label Space Dimension Reduction (LSDR) methods encode high-dimensional labels into lower-dimensional representations.
• Existing LSDR methods often require separate encoding functions, increasing complexity.

Purpose of the Study:

• To propose a novel, efficient, and feature-aware LSDR method called End-to-End Feature-aware label space Encoding (E²FE).
• To directly learn a code matrix in an end-to-end manner, integrating feature awareness.
• To enhance the recoverability of the label space and predictability of the latent space.

Main Methods:

• E²FE learns a code matrix directly from training instances, optimizing label recoverability and latent space predictability.
• Predictive models are trained to map instance features to code vectors.
• A linear decoding matrix is learned for efficient label vector recovery from predicted code vectors.
• The method supports optional encoding function learning and kernel extensions for nonlinear correlations.

Main Results:

• Theoretical analysis confirms efficient learning of both code and decoding matrices.
• Comprehensive experiments on diverse, large-scale datasets demonstrate consistent performance gains.
• E²FE outperforms existing state-of-the-art LSDR methods.

Conclusions:

• E²FE offers a significant advancement in LSDR for multilabel classification.
• The feature-aware, end-to-end approach provides improved efficiency and accuracy.
• The method is versatile and adaptable to nonlinear data structures.