Nonlinear Asymmetric Multi-Valued Hashing.

Area of Science:

• Computer Science
• Machine Learning
• Data Mining

Background:

• Existing hashing methods use binary codes for efficient large-scale similarity search.
• Binary codes often lack sufficient similarity preservation, leading to suboptimal performance.

Purpose of the Study:

• To propose Nonlinear Asymmetric Multi-Valued Hashing (NAMVH) for improved similarity search.
• To address the limitations of binary codes in preserving data similarity.

Main Methods:

• Employing two distinct non-binary embeddings: a real-valued embedding for queries and a multi-integer embedding for database compression using Binary Sparse Representation (BSR).
• Jointly learning embeddings via an alternative optimization method to solve a mixed integer programming problem, preserving label-based similarity.
• Enabling incremental extension for dynamically evolving databases.

Main Results:

• NAMVH outperforms existing binary hashing methods in search accuracy across seven large-scale datasets.
• The proposed method retains the query and storage efficiency characteristic of binary hashing approaches.
• Demonstrated superior similarity preservation compared to traditional binary hashing techniques.

Conclusions:

• NAMVH offers a more accurate and efficient solution for large-scale semantic search.
• The use of non-binary embeddings significantly enhances similarity preservation capabilities.
• The method provides a robust framework for handling dynamic datasets in similarity search applications.