Reproduce “The Effects of Data Quality on Machine Learning Performance on Tabular Data” on OpenResearch

About this paper

Sedir Mohammed, Lukas Budach, Moritz Feuerpfeil, Nina Ihde, Andrea Nathansen, Nele Noack, Hendrik Patzlaff, Felix Naumann and 1 more

Researchers from the Hasso Plattner Institute systematically investigated how six data quality dimensions affect 19 machine learning algorithms across classification, regression, and clustering tasks. The study quantified that completeness, feature accuracy, and target accuracy critically degrade model performance, while uniqueness and consistent representation have a surprisingly low impact on supervised learning on tabular data.

Key insights

• Completeness, feature accuracy, and target accuracy consistently cause high performance degradation across all machine learning tasks.
• Uniqueness and consistent representation often have a surprisingly low impact on the performance of supervised ML models on tabular data.
• Machine learning models trained on clean data are highly vulnerable to data quality issues present in test or production data, underscoring the need for robust data validation.

Summary by alphaXiv.