A New Vector Indexing Algorithm for Embedded AI Systems
Lists
Only you see this
Abstract
To address the efficiency of degradation and declining recall in traditional high-dimensional vector indexing methods, this paper proposes an innovative Vector Partition Tree (VPT) algorithm that leverages vector-space symmetry. The main contributions of this work are as follows: presenting a novel hierarchical VPT data structure for efficient high-dimensional indexing; designing a new interface methods that enable dynamic tree management; and introducing an adaptive node-expansion model to mitigate sparsity, allowing the VPT to recursively construct hierarchical partition structures that adaptively counteract high-dimensional sparsity. Experiments demonstrate VPT’s significant superiority over k-dimensional (K-D) tree and Locality-Sensitive Hashing (LSH) in sparse high-dimensional scenarios, reducing index construction time by up to 61% and achieving a 90% Top-100 recall rate. However, VPT’s retrieval performance becomes unstable in dense low-dimensional environments due to static partition-boundary effects and leaf-node overload. Future research will focus on partition-boundary optimization, dynamic node splitting, and hybrid storage architectures to enhance the adaptability of dense vectors and overcome memory bottlenecks. This work provides an efficient and precise solution for high-dimensional vector retrieval.
BibTeX
@misc{you2025new,
author = {Shan You and Tomoyuki Nagase and Tsutomu Zeniya and Toshiki Kanamoto},
title = {{A New Vector Indexing Algorithm for Embedded AI Systems}},
year = {2025},
doi = {10.1109/gcce65946.2025.11275128},
}Loading related content...