[Paper Review] Efficient and Effective Table-Centric Table Union Search in Data Lakes

Yongkang Sun, Zhihao Ding+3 more|arXiv (Cornell University)|Mar 18, 2026

Data Quality and Management0 citations

TL;DR

TACTUS introduces a table-centric approach for top-k table union search by first learning table embeddings to capture holistic table-level unionability, then refining with column-level cues for final ranking, achieving better accuracy and speed than column-centric methods.

ABSTRACT

In data lakes, information on the same subject is often fragmented across multiple tables. Table union search aims to find the top-k tables that can be unioned with a query table to extend it with more rows, without relying on metadata or ground-truth labels. Existing methods are mainly column-centric: they focus on modeling column unionability scores using column embeddings, which are then used throughout the search process for column matching, filtering, and aggregation. However, this overlooks holistic table-level semantics, which may result in suboptimal rankings and inefficiencies. We introduce TACTUS, a novel table-centric method for table union search. Unlike prior work that searches from columns to tables, we search in a table-first way and examine columns only in the final step. During offline processing, we directly generate table embeddings for holistic, table-level unionability scoring by designing table-level representation techniques, including positive table pair construction to simulate unionable tables, two-pronged negative table sampling to avoid latent positives and mine hard negatives to enhance representation quality, and attentive table encoding for effective embeddings. During online search, we first develop a table-centric adaptive candidate retrieval method that efficiently selects a compact, high-quality candidate pool by leveraging the distribution of table-level unionability scores induced by table embeddings. We then inspect columns only within this compact candidate set and design a dual-evidence reranking technique that integrates table-level and column-level scores to refine the final top-k results. Extensive experiments on real-world datasets show that TACTUS significantly improves result quality while being much faster than existing methods in both offline and online processing, often by an order of magnitude.

Motivation & Objective

• Motivate the need to search for unionable tables without relying on metadata in data lakes.
• Propose a table-centric framework to capture holistic table-level semantics for unionability.
• Develop offline table embedding learning with self-supervised signals to distinguish unionable vs non-unionable tables.
• Design an online retrieval and reranking pipeline that combines table-level and column-level evidence for top-k results.

Proposed method

• Produce a table embedding Ti for each table Ti in the data lake to encode table-level unionability semantics.
• Offline: construct positive table pairs and two-pronged negative sampling to train an attentive table encoder via a table-level contrastive loss.
• Online: perform table-centric adaptive candidate retrieval using table embeddings to form a compact candidate set S, then rerank with a dual-evidence score combining mu_T and mu_A.
• Mu_T is a table-level unionability score from embeddings; mu_A is a column-alignment score obtained offline via FastText-based features, combined into mu for final ranking.

Experimental results

Research questions

• RQ1 How can a table-centric representation capture holistic table-level unionability without relying on table metadata?
• RQ2 Can a two-stage process (table-level retrieval followed by column-level refinement) outperform column-centric approaches in both accuracy and efficiency?
• RQ3 What self-supervised strategies best approximate unionable table pairs to train table embeddings?
• RQ4 Does adaptive candidate retrieval yield a compact, high-coverage set that speeds online search without sacrificing top-k quality?

Key findings

• TACTUS consistently achieves superior effectiveness compared to existing methods in the reported experiments.
• The table-centric pipeline yields significant speedups over prior approaches in both offline and online processing, often by an order of magnitude.
• A single table embedding per table enables efficient candidate retrieval and early pruning of non-relevant tables.
• Dual-evidence reranking effectively combines table-level embeddings with column-level alignment to refine top-k results.
• Offline, self-supervised table embeddings with positive table pair construction and hard-negative sampling improve the quality of unionability estimation.