Title: Scalable Vector Analytics: A Story of Twists and Turns

 

Speaker: Professor Themis Palpanas(University Paris Cite and IUF)

 

Abstract: Similarity search in high-dimensional data spaces was a relevant and challenging data management problem in the early 1970s, when the first solutions to this problem were proposed. Today, fifty years later, we can safely say that the exact same problem is more relevant (from Time
Series Management Systems to Vector Databases) and challenging than ever. Very large amounts of high-dimensional data are now omnipresent (ranging from traditional multidimensional data to time series and deep embeddings), and the performance requirements (i.e., response-time and accuracy) of a variety of applications that need to process and analyze these data have become very stringent and demanding. In these past fifty years, high-dimensional similarity search has been studied in its many flavors. Similarity search algorithms for exact and approximate, one-off and progressive query answering. Approximate algorithms with and without (deterministic or probabilistic) quality guarantees. Solutions for on-disk and in-memory data, static and streaming data. Approaches based on multidimensional space-partitioning and metric trees, random projections and locality-sensitive hashing (LSH), product quantization (PQ) and inverted files, k-nearest neighbor graphs and optimized linear scans. Surprisingly, the work on data-series (or time-series) similarity search has recently been shown to achieve the state-of-the-art performance for several variations of the problem, on both time-series and general high-dimensional vector data. In this talk, we will touch upon the different aspects of this interesting story, present some of the state-of-the-art solutions, and discuss open research directions.