Prof Dan Olteanu : 'In-Database Factorized Learning'

Abstract: In this talk I will overview work on compilation of join queries into lossless factorized representations. The primary motivation for this compilation is to avoid redundancy in the representation of query results in relational databases. The relationship between the standard tabular representations of relations and equivalent factorized representations is on a par with the relationship between propositional formulas in disjunctive normal form and equivalent circuits. For any join query, we give asymptotically tight bounds on the size of its factorized results and on the time to compute them. We show that these factorized results can be exponentially more succinct than their equivalent tabular representations.

I will also discuss an application of factorized joins to learning regression models over databases.  On-going joint work with LogicBlox collaborators shows that in-database factorized learning can speed up real analytical workloads in the retailer domain by several orders of magnitude over state-of-the-art systems. 

This work is based on long-standing collaboration with Maximilian Schleich and Jakub Zavodny and more recent collaboration with Hung Ngo and XuanLong Nguyen. 

Bio: Dan Olteanu is a professor of Computer Science at Oxford, Alan Turing Institute faculty fellow, and computer scientist at LogicBlox. He has also taught at the universities of California Berkeley, Munich, Saarland, and Heidelberg. He received his PhD in Computer Science from University in Munich in 2005. His research interests are in databases and adjacent areas. Dan contributed to XML query processing, incomplete information and probabilistic databases, and more recently to factorized databases and the commercial LogicBlox system. He co-authored the book "Probabilistic Databases" (2011). He has served as associate editor for PVLDB'13 and IEEE TKDE, track chair for ICDE'15, group leader for SIGMOD'15, and vice chair for SIGMOD'17. His current research is supported by an ERC consolidator grant and awards from Amazon, Google, and LogicBlox.