Proximity-based frameworks for generating embeddings from multi-output data.

Area of Science:

• Machine Learning
• Data Science
• Natural Language Processing

Background:

• Dimensionality reduction (DR) is crucial for handling high-dimensional data.
• Existing methods often struggle with multi-output and multilabel data structures.
• Spectral embedding techniques offer powerful approaches for DR.

Purpose of the Study:

• To propose novel frameworks for supervised (SDR) and semi-supervised (SSDR) dimensionality reduction tailored for multi-output data.
• To generalize existing SDR techniques for multilabel classification tasks.
• To provide a unified representation of spectral DR methods and introduce efficient SSDR strategies.

Main Methods:

• Developed two frameworks: MESD (sample duplication for single-to-multilabel SDR) and MOPE (simultaneous feature-label proximity for multilabel SDR).
• Proposed diverse schemes for label-based proximity calculation and feature-label information fusion within MOPE.
• Introduced a general framework for SSDR by combining SDR and UDR models, and a cost-reduction procedure using target relation features.

Main Results:

• Demonstrated the effectiveness of the proposed MESD and MOPE frameworks on document collections for multilabel text categorization.
• Showcased the utility of the unified template for various spectral DR methods (UDR, SDR, SSDR).
• Validated the efficiency of the SSDR framework and the computational cost reduction technique.

Conclusions:

• The proposed frameworks offer flexible and effective solutions for supervised and semi-supervised dimensionality reduction in multilabel classification.
• The unified representation and SSDR framework advance the field of spectral methods for complex data.
• Experimental results confirm the practical applicability and performance gains in natural language processing tasks.