Reproduce “Retrieve, Merge, Predict: Augmenting Tables with Data Lakes” on OpenResearch

About this paper

Riccardo Cappuzzo, Aimee Coelho, Felix Lefebvre, Paolo Papotti, Gael Varoquaux

Researchers systematically evaluate a 'Retrieve, Merge, Predict' pipeline for augmenting tabular data from unstructured data lakes, finding that tree-based models like CatBoost are robust and simpler retrieval/aggregation methods often suffice for performance gains with lower computational cost. The study also introduces YADL, a novel semi-synthetic data lake, to foster reproducible benchmarking in this domain.

Key insights

• Tree-based models like CatBoost are highly effective and robust for machine learning on tables augmented from data lakes, handling missing values and high-cardinality features well.
• Simple, metric-based retrieval methods (e.g., Exact Matching, Hybrid MinHash) often provide competitive or superior prediction performance compared to more complex, resource-intensive approaches like Starmie.
• Performance gains from augmenting tables tend to plateau quickly, suggesting diminishing returns from integrating a large number of additional candidate tables or using overly complex aggregation methods.

Summary by alphaXiv.