Scalable Supervised Asymmetric Hashing With Semantic and Latent Factor Embedding

Area of Science:

• Computer Science
• Machine Learning
• Data Mining

Background:

• Compact hash code learning is crucial for fast similarity search due to reduced storage and efficient querying.
• Learning discriminative binary codes that preserve full pairwise similarities in high-dimensional data remains a significant challenge.

Purpose of the Study:

• To propose a novel scalable supervised asymmetric hashing (SSAH) method to address the limitations of existing hash code learning techniques.
• To enhance the preservation of full pairwise similarities for improved search performance.

Main Methods:

• Developed a scalable supervised asymmetric hashing (SSAH) method utilizing maximum asymmetric inner product of two distinct non-binary embeddings.
• Integrated supervised label information and refined latent feature embeddings to construct high-quality hashing functions.
• Employed an alternating optimization scheme with guaranteed convergence to solve the discrete optimization problem.

Main Results:

• The SSAH method effectively approximates the full-pairwise similarity matrix, avoiding information loss.
• Simultaneously minimized regression loss for semantic label alignment and encoding loss for refined latent features.
• Demonstrated superior performance on diverse benchmark datasets compared to recently proposed hashing algorithms.

Conclusions:

• The proposed SSAH method offers a significant advancement in learning discriminative binary codes for similarity search.
• SSAH effectively balances the trade-off between storage efficiency, query speed, and accuracy in high-dimensional data.
• The method provides a robust and scalable solution for practical similarity search applications.