Applicability of Linear Dimensionality Reduction Methods to Approximate Nearest Neighbor Search with Out-of-Distribution Queries
Received: 09 Nov 2026 Accepted: 09 Nov 2026
Published: 2026, vol. 30, issue 4, pp. 324–327
Abstract
We study linear dimensionality reduction and quantization for approximate nearest neighbor search (ANNS) when the distributions of indexed vectors and queries differ. We show that discarding components that are small for indexed vectors may cause significant dot-product errors for out-of-distribution (OOD) queries. We propose a sequential decomposition into orthogonal subspaces to account for components important to OOD queries. On the 200-dimensional Text-to-Image dataset, the proposed approach reduces vector storage by a factor of 1.67 and accelerates search by a factor of 1.14 compared with PCA-based quantization.
Keywords: approximate nearest neighbor search, dimensionality reduction, out-of-distribution queries, linear transformations, quantization.
BibTeX
@article{IS-Aliseychik-Snegova2026,
author = {Aliseychik, Pavel A. and Snegova, Elena A.},
title = {{Applicability of Linear Dimensionality Reduction Methods to Approximate Nearest Neighbor Search with Out-of-Distribution Queries}},
journal = {Intelligent Systems. Theory and Applications},
year = {2026},
volume = {30},
number = {4},
pages = {324--327},
}
AMSBIB
\Bibitem{IS-Aliseychik-Snegova2026}
\by P.\,A.~Aliseychik, E.\,A.~Snegova
\paper Applicability of Linear Dimensionality Reduction Methods to Approximate Nearest Neighbor Search with Out-of-Distribution Queries
\jour Intelligent Systems. Theory and Applications
\yr 2026
\vol 30
\issue 4
\pages 324--327
\lang In Russian
RU