Human-Centered Exploration of Table Unionability

Table union search (TUS) identifies tables that can be meaningfully combined with a given query table and is a core task in data discovery over data lakes. Yet, what it means for two tables to be "unionable" is inherently ambiguous: domain experts may disagree even on seemingly simple cases, and existing benchmarks collapse this disagreement into binary labels, omitting the behavioral context behind human decisions. We take a human-centered view of table unionability and study how humans, traditional TUS methods, and large language models (LLMs) interact on this task. We introduce TUNE (Table UNionability with human Evaluation), a benchmark of 464 expert judgments over 26 table pairs that records binary decisions, confidence scores, decision times, interaction traces, textual explanations, and post-survey reflections. Using TUNE, we (i) characterize human performance, overconfidence, and metacognitive quality (calibration and resolution); (ii) benchmark state-of-the-art TUS methods (Starmie, SANTOS, D3L), revealing complementary strengths and systematic misalignment with human judgments; and (iii) evaluate four experimental scenarios that combine human behavioral signals and TUS features using classical ML models and LLMs. The best configuration we experimented with reaches 84% accuracy, improving over both human majority vote and the strong standalone TUS method, while LLMs act as useful second opinions but are sensitive to conflicting signals. Overall, our results suggest that unionability labels reflect a structured yet imperfect human decision process and that hybrid human-model (may it be traditional classifiers or LLMs) pipelines provide more reliable and interpretable unionability assessments.