MosaicJoin: Compact Semantic Sketches for Value-Level Join Discovery

MosaicJoin introduces a training-free method for semantic join discovery at the value level. It uses compact sketches to approximate whether two columns can be joined, avoiding exhaustive comparisons for high-cardinality columns. At query time, candidate scoring is bounded by sketch size, while a query-subsampling operator reduces search cost with provable accuracy guarantees. The abstract reports that MosaicJoin outperforms previously published methods across all benchmarks and runs up to 66 times faster than other value-level approaches. It supports query columns with up to 57K values and data-lake columns with up to 1M values.