Joinable Table Search in Data Lakes

Defines the joinable table search problem as a top-k overlap set similarity query. Introduces a cost-aware algorithm that mixes prefix-filtering with position-list scans to scale to lakes with millions of columns. The reference exact baseline for nearly all follow-on work.

Tackles the variable-cardinality problem: when column sizes range from hundreds to millions, single LSH indexes degrade badly. The ensemble partitions columns by size and tunes each partition's LSH parameters to approximate the containment measure efficiently.

Embeds textual cell values as high-dimensional vectors using pre-trained word embeddings and defines joinability via similarity predicates on those vectors. Introduces a pivot-based block-and-verify index and a partitioning scheme for out-of-core data lakes.

Fine-tunes a PLM to embed entire columns (rather than individual cells) into a fixed-length vector, unifying equi- and semantic joins under one retrieval framework. Couples the encoder with HNSW for approximate nearest-neighbor search; retrieval becomes logarithmic in repository size.

Uses pre-trained WebTable (FastText) embeddings rather than task-specific fine-tuning, indexed with HNSW. Aimed at cloud warehouse settings; a useful counterpoint to DeepJoin for an embeddings-without-training baseline.

Targets joinability discovery in data products — curated collections of tabular datasets — rather than open data lakes. Combines multiple column-pair similarity signals through a self-supervised Graph Neural Network (RGCN), so similarity propagates across the column-relatedness graph and indirect join relationships emerge. Uses automated negative-example generation to sharpen precision, and reports up to 14% F1/AUC gains over DeepJoin and other baselines without per-metric thresholds.

Models a table corpus as a hypergraph in which columns are nodes and hyperedges encode joinability — including LLM-augmented inter-table hyperedges. Joinable table discovery is recast as link prediction over the hypergraph, with a maximum-spanning-tree reranker for coherence. Reports +21% Precision@15 / +17% Recall@15 over the best prior baselines.

Introduces a hash-based index supporting n-ary join discovery via space-efficient "super keys" that combine multiple columns. Demonstrates significant gains over applying unary methods combinatorially.

Tackles the gap between single-key embedding methods (e.g., DeepJoin) and hash-based n-ary discovery (e.g., MATE): combining keys before embedding loses semantics, while embedding each key independently breaks correlations across keys. PolyJoin learns joint representations over sets of key columns so that semantic multi-key joinability — e.g., (first_name, last_name) or (city, state) — can be retrieved with approximate nearest-neighbor search.

Identifies promising joinable row pairs as input/output examples and uses program-synthesis-style search to find a transformation that aligns the columns. Scales to tens of thousands of rows at interactive speed.

A by-example formulation of join discovery: users provide a small set of (row, expected match) examples and the system retrieves and ranks candidate joinable tables consistent with them. Complements both value-overlap and embedding approaches when the user's notion of "joinable" is implicit and best conveyed through samples.

Defines context-aware column joinability and proposes a multi-criteria scoring approach (TOPJoin) that combines value, header, and contextual signals. Includes an experimental comparison against JOSIE, DeepJoin, and WarpGate.

A system that combines, n-ary join search with union search and explores optimization of discovery operators.

A large-scale, labeled benchmark covering both joinable and unionable table search across multiple domains. Includes baseline comparisons and standardized metrics — the natural starting point for project evaluations.

An end-to-end study of table-augmentation pipelines: retrieve joinable tables, merge them, train a model. Introduces YADL (Yet Another Data Lake) for benchmarking and reports several surprising findings — e.g., Jaccard containment is a strong retrieval criterion, and tree-based learners are more robust to noisy merges than deep models.