TabArena - A Living Benchmark for Machine Learning on Tabular Data

Learn about TabArena, the first continuously maintained living benchmarking system for machine learning on tabular data, in this 46-minute seminar presentation. Discover how this innovative platform addresses the limitations of static benchmarks by providing ongoing updates when flaws are discovered, model versions change, or new models emerge. Explore the comprehensive benchmarking study that compares various machine learning approaches including gradient-boosted trees, deep learning methods, and foundation models across a curated collection of representative datasets. Understand the critical influence of validation methods and hyperparameter ensembling on model performance evaluation, and examine findings showing that while gradient-boosted trees remain competitive on practical datasets, deep learning methods have achieved parity under larger time budgets with ensembling, and foundation models particularly excel on smaller datasets. Gain insights into how model ensembles advance the state-of-the-art in tabular machine learning and investigate individual model contributions to overall performance. Access the public leaderboard, reproducible code, and maintenance protocols that make TabArena a living system for the tabular ML community.