报告主题：MIRAGE: A Fast and Scalable Graph-Based Index for Vector Search

报告人： M. Tamer Özsu, University of Waterloo

报告时间：2026年1月21日 上午 09:30

报告地点：北京大学王选计算机研究所106报告厅

Abstract：Approximate Nearest Neighbor Search (ANNS) over high-dimensional vector spaces is a foundational problem in numerous domains, including information retrieval, recommender systems, and, increasingly, large language models (LLMs), where methods such as Retrieval-Augmented Generation (RAG) depend critically on efficient vector search to integrate external knowledge. Among the leading approaches for ANNS, graph-based vector indexes have shown superior performance in balancing retrieval accuracy and latency. However, these structures often incur significant computational costs during index construction, suffer from complex parameter tuning requirements, and exhibit limited scalability in high-throughput settings.

This talk surveys the landscape of graph-based indexing methods, outlining their algorithmic foundations, empirical strengths, and practical limitations. I will then present MIRAGE, a new indexing framework that addresses several key challenges in current methods. MIRAGE is designed to reduce construction overhead while maintaining high retrieval quality, and it introduces novel strategies for improving search robustness and memory efficiency. I will discuss the design principles behind MIRAGE, its integration into large-scale systems, and results from empirical evaluations against leading ANNS baselines. The talk concludes with open challenges and future directions in scalable, high-performance vector search.（Joint work with Sairaj Voruganti.）

Bio: M. Tamer Özsu is a University Professor in the David R. Cheriton School of Computer Science at the University of Waterloo. His research focuses on the data engineering aspects of data science, with particular emphasis on distributed data management and the management of non-conventional data types. Dr. Özsu is a Fellow of the Royal Society of Canada, the American Association for the Advancement of Science, the Science Academy of Türkiye, the Asia-Pacific Artificial Intelligence Association, and the Balsillie School of International Affairs. He is also a Life Fellow of both the ACM and IEEE. His distinctions include the University of Waterloo Award of Excellence in Graduate Student Supervision (2025), the ACM Presidential Award (2024), the IEEE Technical Committee on Data Engineering Education Award (2024), the IEEE Innovation in Societal Infrastructure Award (2022), the CS-Can/Info-Can Lifetime Achievement Award (2018), the ACM SIGMOD Test-of-Time Award (2015), the ACM SIGMOD Contributions Award (2006), and The Ohio State University College of Engineering Distinguished Alumnus Award (2008). His publications have received four Best Paper Awards and one Honourable Mention.

Dr. Özsu was the Founding Editor-in-Chief of ACM Books (2014–2020) and the Founding Series Editor of Synthesis Lectures on Data Management (2009–2014). He currently serves on the editorial boards of three journals and one book series.