TupleRank and Implicit Relationship Discovery in Relational Databases

Google's successful PageRank brings to the Web an order that well re ects the relative importance of Web pages. Inspired by PageRank, we propose a similar scheme called TupleRank for ranking tuples in a relational database. Database tuples naturally relate to each other through referential integrity constraints declared in the schema. However, such constraints cannot capture more general relationships such as similarity. Furthermore, relationships determined statically from the database schema do not re ect actual query patterns that arise at runtime. To address these de ciencies of static TupleRank, we introduce the notion of query-driven TupleRank, which is based on a link structure that is dynamically constructed from a workload. Speci cally, database tuples are considered to be related if they are joined together by a query result tuple. The main challenge in supporting query-driven TupleRank is how to record multiway relations among tuples eÆciently for a large workload. We develop techniques to compute query-driven TupleRank accurately and eÆciently with low space requirement that is independent of the workload size. We further augment query-driven TupleRank so that it can better utilize the access frequency information collected from the workload. Preliminary experiment results demonstrate that TupleRank is both informative and intuitive, and they con rm the advantages of query-driven TupleRank over static TupleRank.