July 2, 2026
MosaicJoin: Compact Semantic Sketches for Value-Level Join Discovery Accepted to PVLDB
News2026-07-02
MosaicJoin: Compact Semantic Sketches for Value-Level Join Discovery, co-authored by Grace Fan, Majid Daliri, and Eden Wu, has been accepted by PVLDB.
"MosaicJoin: Compact Semantic Sketches for Value-Level Join Discovery" was accepted to PVLDB! Congrats to Grace Fan, Eden Wu, and Majid Daliri.
The paper is about semantic join discovery, finding columns in a data lake that can be joined even when the values don't match exactly (think "NYC" vs "New York City"). Existing methods force you to choose: compare values directly to get accuracy but no scalability, or compress each column into a single embedding to get speed but lose the value-level detail that actually tells you whether a join works. MosaicJoin gets both, using a sketching approach that estimates joinability without comparing every value pair, plus a subsampling trick with accuracy guarantees for large query columns. It beats prior methods on every benchmark we tried and is up to 66x faster, with no training needed.
The paper is about semantic join discovery, finding columns in a data lake that can be joined even when the values don't match exactly (think "NYC" vs "New York City"). Existing methods force you to choose: compare values directly to get accuracy but no scalability, or compress each column into a single embedding to get speed but lose the value-level detail that actually tells you whether a join works. MosaicJoin gets both, using a sketching approach that estimates joinability without comparing every value pair, plus a subsampling trick with accuracy guarantees for large query columns. It beats prior methods on every benchmark we tried and is up to 66x faster, with no training needed.