In Defense of Locality-Sensitive Hashing.

Area of Science:

• Computer Science
• Information Retrieval
• Machine Learning

Background:

• Hashing-based semantic similarity search is crucial for large-scale content retrieval systems.
• Current supervised hashing methods involve computationally intensive two-step strategies.
• Locality-sensitive hashing (LSH) has been overlooked for generating hash codes due to performance limitations.

Purpose of the Study:

• To introduce a novel hashing framework, locality-sensitive two-step hashing (LS-TSH).
• To leverage LSH for transforming semantic labels into binary codes efficiently.
• To demonstrate the effectiveness of LS-TSH in semantic similarity search.

Main Methods:

• Developed LS-TSH, a two-step hashing framework utilizing LSH.
• Generated binary codes by transforming semantic labels via LSH.
• Avoided complex binary optimization problems common in other methods.

Main Results:

• LS-TSH achieves comparable retrieval accuracy to state-of-the-art methods.
• LS-TSH demonstrates training speeds two to three orders of magnitude faster.
• Theoretically, LS-TSH preserves label-based semantic similarity and offers sublinear query complexity.

Conclusions:

• LS-TSH presents an effective and efficient approach for semantic similarity search.
• The proposed method significantly reduces computational cost during training.
• LS-TSH redefines the utility of LSH in generating effective hash codes for retrieval systems.