Multilabel Prediction via Cross-View Search.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Embedding methods show promise in multilabel prediction by capturing label dependencies.
• Existing methods often neglect input-output correlations, leading to misaligned embeddings and reduced performance.
• This misalignment hinders the effectiveness of learned representations in complex prediction tasks.

Purpose of the Study:

• To develop a novel formulation for multilabel learning that addresses the input-output correlation issue.
• To propose a method that jointly learns embeddings and mappings, preserving semantic similarity.
• To enhance prediction efficiency through optimized search strategies in learned embedding spaces.

Main Methods:

• Introduced Co-Embedding (CoE), a cross-view learning framework that jointly learns a common semantic subspace and view-specific mappings.
• Preserved semantic similarity structure among embeddings to ensure related embeddings share similar labels.
• Developed Co-Hashing (CoH), a hashing-based extension of CoE for compact binary representations and efficient prediction.

Main Results:

• CoE effectively explores input-output correlations, leading to better-aligned embeddings.
• Cross-view k-nearest neighbor (kNN) search in CoE significantly reduces computational costs compared to traditional decoding.
• CoH achieves superior prediction accuracy and efficiency on various real-world datasets, outperforming state-of-the-art methods.

Conclusions:

• The proposed CoE and CoH methods offer a robust approach to multilabel learning by leveraging cross-view correlations.
• These methods significantly improve both prediction accuracy and computational efficiency.
• The findings highlight the importance of aligning input and output representations for effective multilabel prediction.