Robust hashing with local models for approximate similarity search

Area of Science:

• Computer Science
• Data Science
• Machine Learning

Background:

• High-dimensional data similarity search is crucial for many applications.
• The curse of dimensionality degrades the performance of traditional indexing structures.
• Existing hashing methods like locality-sensitive hashing (LSH) offer efficiency but often ignore data structure.

Purpose of the Study:

• To propose a novel hashing method, Robust Hashing with Local Models (RHLM), for efficient similarity search in high-dimensional data.
• To leverage local structural information within the data to generate robust hash codes.
• To improve upon the limitations of existing randomized hashing techniques.

Main Methods:

• RHLM learns local hashing models for individual data points, predicting hash codes for neighbors.
• Global alignment of local models ensures optimal hash code assignment.
• Utilizes l2,1-norm minimization for robust hash function learning and maps database points to hash codes.

Main Results:

• RHLM effectively utilizes local structural information to create accurate hash codes.
• The method demonstrates superior search quality compared to state-of-the-art techniques.
• Experimental results on real-life datasets confirm RHLM's enhanced efficiency.

Conclusions:

• RHLM offers a robust and efficient approach to approximate similarity search in high-dimensional spaces.
• The integration of local data structure significantly improves hashing performance.
• RHLM represents a promising advancement over traditional and randomized hashing methods.