跳到主要导航 跳到搜索 跳到主要内容

Tree and Graph Based Two-Stages Routing for Approximate Nearest Neighbor Search

  • Jiannan Li
  • , Zhenyu Zhang
  • , Xiaoling Wang*
  • , Haoyang Li
  • *此作品的通讯作者
  • East China Normal University
  • Montverde Academy Shanghai

科研成果: 书/报告/会议事项章节会议稿件同行评审

摘要

With the expansion of extensive datasets of high-dimensional vectors, the approximate nearest neighbor search (ANNS) has become increasingly significant in data mining. Among the prevailing ANNS methods, graph-based methods achieve high recall at the cost of excessive memory usage, whereas tree-based methods are memory-efficient but suffer from low recall and increased search complexity. In this paper, we propose a hybrid index based on Ball tree and Hierarchical Navigable Small World(HNSW). Ball-HNSW employs a two-stages hybrid framework to accelerate the identification of approximate nearest neighbors. For datasets with different distributions, we utilize reinforcement learning to optimize parameter combinations by simulating decisions during the search process, thereby improving search results. Experiments confirm that Ball-HNSW outperforms other state-of-the-art methods on multiple datasets.

源语言英语
主期刊名Web and Big Data - 8th International Joint Conference, APWeb-WAIM 2024, Proceedings
编辑Wenjie Zhang, Zhengyi Yang, Xiaoyang Wang, Anthony Tung, Zhonglong Zheng, Hongjie Guo
出版商Springer Science and Business Media Deutschland GmbH
376-390
页数15
ISBN(印刷版)9789819772377
DOI
出版状态已出版 - 2024
活动8th Asia-Pacific Web and Web-Age Information Management Joint International Conference on Web and Big Data, APWeb-WAIM 2024 - Jinhua, 中国
期限: 30 8月 20241 9月 2024

出版系列

姓名Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
14963 LNCS
ISSN(印刷版)0302-9743
ISSN(电子版)1611-3349

会议

会议8th Asia-Pacific Web and Web-Age Information Management Joint International Conference on Web and Big Data, APWeb-WAIM 2024
国家/地区中国
Jinhua
时期30/08/241/09/24

指纹

探究 'Tree and Graph Based Two-Stages Routing for Approximate Nearest Neighbor Search' 的科研主题。它们共同构成独一无二的指纹。

引用此