EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Cost-Aware Decoupled Optimization for Graph-Based Vector Index Tuning With Adaptive Budget Allocation

Hyeonbyeong Lee, Sangho Song, Donghyeon Cha, Dojin Choi, Kyoungsoo Bok, Jaesoo Yoo

IEEE Access · 2026 · 14 · 69881-69895

DOI: 10.1109/access.2026.3689317 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Hyeonbyeong Lee und 5 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Glocal University Project of Mokpo National University, in 2026,, Innovative Human Resource Development for Local Intellectualization Program through the Institute of Information and Communications Technology Planning and Eva…, Korean Government [Ministry of Science and ICT (MSIT)], 50%.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Graph-based vector indexes expose a hierarchical parameter structure: build-time parameters determine the index topology and are fixed after construction, while search-time parameters control query traversal on the fixed graph. Existing auto-tuners treat these as a flat joint space, ignoring the cost asymmetry—build cost can exceed search cost by orders of magnitude—and the hierarchical dependency whereby an underbuilt topology caps the recall-throughput frontier regardless of search-time tuning. We propose Decoupled Optimization Framework (DOF)-TC and ADOF-TC, staged optimization frameworks that decouple build-stage exploration from search-stage refinement. The build stage screens candidates via target-conditioned proxy scoring that evaluates throughput at the deployment recall target, forwarding Pareto-optimal topologies to a search stage equipped with an early-target efficiency bonus. ADOF-TC further adapts the budget split online by measuring build/search cost ratios in situ. To determine when staged optimization is preferable to monolithic tuning, we introduce the Build-to-Budget Ratio (BBR)—the fraction of the total optimization budget consumed by a single index build. Experiments on six datasets spanning the BBR spectrum (0.07–0.72) show that staged methods achieve + 80–110% throughput in the low-BBR regime, while monolithic methods dominate as BBR increases, with a crossover between BBR 0.17 and 0.33.

Förderung

  • Glocal University Project of Mokpo National University, in 2026,
  • Innovative Human Resource Development for Local Intellectualization Program through the Institute of Information and Communications Technology Planning and Eva…
  • Korean Government [Ministry of Science and ICT (MSIT)], 50%

Zitieren

APA

Lee, H., Song, S., Cha, D., Choi, D., Bok, K., & Yoo, J. (2026). Cost-Aware Decoupled Optimization for Graph-Based Vector Index Tuning With Adaptive Budget Allocation. IEEE Access, 14, 69881-69895. https://doi.org/10.1109/access.2026.3689317

Vancouver

Lee H, Song S, Cha D, Choi D, Bok K, Yoo J. Cost-Aware Decoupled Optimization for Graph-Based Vector Index Tuning With Adaptive Budget Allocation. IEEE Access. 2026;14:69881-69895. doi:10.1109/access.2026.3689317

BibTeX
@article{Lee2026,
  author = {Lee, Hyeonbyeong and Song, Sangho and Cha, Donghyeon and Choi, Dojin and Bok, Kyoungsoo and Yoo, Jaesoo},
  title = {Cost-Aware Decoupled Optimization for Graph-Based Vector Index Tuning With Adaptive Budget Allocation},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {69881--69895},
  doi = {10.1109/access.2026.3689317}
}
Kostenlos starten →