2023年7月26日学术报告——Hua Lu

报告题目：Learned Indexes for Spatial Data

报告人：Hua Lu

时间：2023年7月26日上午9点30

地点：计算机学科楼338

报告人简介：

Hua Lu (https://luhua.ruc.dk/) is a Professor of Computer Science, the Department of People and Technology, Roskilde University, Denmark. From 2007 to 2020 he worked in the Department of Computer Science, Aalborg University, Denmark. He received the BSc and MSc degrees from Peking University, China, and the PhD degree in Computer Science from National University of Singapore. His research interests span databases, data science, spatial data, and geographic information systems. He has published more than 150 peer-reviewed papers, with 60+ in CCF-A publication venues (e.g., SIGMOD, PVLDB, ICDE, WWW and TKDE). He received the Best Student Paper Runner-up Award at ADC 2022, Best Paper Nomination at SSTD 2021, and the Best Vision Paper Award at SSTD 2019. He has served as PC cochair or vice-chair for ISA 2011, MUE 2011, MDM 2012, NDBC 2019 and IEEE BigData 2022, demo chair for SSDBM 2014, and PhD forum cochair for MDM 2016 and 2022. He has also served on the program committee of many conferences including VLDB, ICDE, WWW, KDD, CIKM, SSTD, ACM SIGSPATIAL GIS and others. He is a senior member of IEEE.

报告内容：

R-tree is a conventional and popular spatial index for supporting many different types of queries over spatial data. However, R-tree may incur large storage consumption and high IO cost in many scenarios. Inspired by the original recursive model index (RMI) that replaces B-tree with staged machine learning models, learned spatial indexes have been proposed to replace R-tree. This talk will present two recent learned indexes for spatial point data. The first index Z-order Model (ZM) simply employs the Z-order curve to transform 2D points to 1D values, and subsequently indexes such 1D values using an RMI-type learned index. The ZM index supports spatial range queries. The second, more sophisticated index LISA employs machine learning models, through several steps, to generate searchable data layout in disk pages for an arbitrary point dataset. Initially, LISA uses a mapping function to transform multi-dimensional points into 1D mapped values. Subsequently, LISA learns a prediction function to partition the mapped value space into shards. Next, LISA learns a series of local models that organize shards into disk pages. LISA supports both range queries and KNN queries efficiently. The latter is enabled by a lattice regression model that converts a KNN query into a number of range queries. After ZM and LISA, the talk will end with brief discussions of learned indexes for polygon data and trajectory data.