A General Framework for Linear Distance Preserving Hashing.

Area of Science:

• Computer Science
• Machine Learning
• Data Mining

Background:

• Binary hashing is crucial for efficient approximate nearest neighbor (ANN) search.
• Existing methods often struggle with preserving data distances in Hamming space.
• Compact data representation in Hamming space enables fast computations.

Purpose of the Study:

• To propose a generic binary hashing framework.
• To introduce a novel linear pairwise distance preserving objective.
• To develop three distinct hashing methods (pseudo-supervised, unsupervised, supervised) within this framework.

Main Methods:

• A generic hashing framework with a linear pairwise distance preserving objective and pointwise constraint.
• Instantiating the framework using pseudo-supervised, unsupervised, and supervised learning paradigms.
• Improving the framework by localizing the distance preserving objective.

Main Results:

• The pseudo-supervised method consistently outperforms state-of-the-art unsupervised hashing methods.
• The unsupervised and supervised methods achieve competitive performance against existing algorithms.
• The framework demonstrates effectiveness on four large-scale benchmark datasets.

Conclusions:

• The proposed generic hashing framework effectively addresses the approximate nearest neighbor search problem.
• The developed hashing methods offer improved performance and efficiency.
• This work contributes novel approaches to binary hashing for enhanced data retrieval.