Hybrid Scalar-Product Compression (HSPC): Optimizing Quantization in Vector Retrieval via Hot-Cold Separation
IEEE Access · 2026 · 14 · 107479-107491
DOI: 10.1109/access.2026.3713412 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Ruijuan Wu und 1 weiteren Autor:innen. Die Arbeit nennt 29 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Colleges Nature Science Research Major Project of Anhui Province, Science Research Team Project, Master’s Supervisor Project.
Abstract (Originalsprache)
As the trade-off between storage overhead and query performance in vector retrieval systems becomes increasingly prominent, this paper proposes a hybrid compression algorithm, Hybrid Scalar-Product Compression (HSPC), which integrates scalar and product quantization, building upon the VStream hot-cold separation mechanism. HSPC assigns differentiated quantization compression ratios based on the hot-cold attributes of the data: hot data prioritize low latency and high retrieval accuracy, while cold data focus on improving compression rates while maintaining accuracy. Experiments were conducted using the SIFT1M dataset to compare the proposed HSPC with the original system’s Gorilla compression algorithm, evaluating performance across three dimensions: memory usage, query latency, and recall rate. The experimental results demonstrate that when HSPC employs scalar quantization for hot data and a configuration of scalar quantization as the primary method supplemented by product quantization for cold data, HSPC achieves significant storage compression while maintaining high retrieval accuracy. Compared to the Gorilla algorithm, HSPC exhibits a slight decrease in recall rate but achieves over 70% improvement in storage efficiency. Query latency increases from 7.8 ms to 10.04 ms, yet remains within an acceptable low-latency range (10 ms). When compared to IVF-PQ, the hybrid compression approach increases recall by 25% at the cost of approximately 100 MB of additional memory. HSPC fully leverages the access characteristics of hot and cold data in vector retrieval scenarios, achieving an effective balance among compression ratio, query efficiency, and retrieval accuracy, thereby providing a viable technical path for storage optimization in vector retrieval systems.
Förderung
- Colleges Nature Science Research Major Project of Anhui Province
- Science Research Team Project
- Master’s Supervisor Project
- Introducing Talent Research Initiation Project of Anhui Xinhua University
Zitieren
APA
Wu, R., & Lv, T. (2026). Hybrid Scalar-Product Compression (HSPC): Optimizing Quantization in Vector Retrieval via Hot-Cold Separation. IEEE Access, 14, 107479-107491. https://doi.org/10.1109/access.2026.3713412
Vancouver
Wu R, Lv T. Hybrid Scalar-Product Compression (HSPC): Optimizing Quantization in Vector Retrieval via Hot-Cold Separation. IEEE Access. 2026;14:107479-107491. doi:10.1109/access.2026.3713412
BibTeX
@article{Wu2026,
author = {Wu, Ruijuan and Lv, Teng},
title = {Hybrid Scalar-Product Compression (HSPC): Optimizing Quantization in Vector Retrieval via Hot-Cold Separation},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {107479--107491},
doi = {10.1109/access.2026.3713412}
}
EUVIMED