The effects of data quality on machine learning performance on tabular data

2022-07-29

Information Systems

AI Summary

This study systematically investigates the mechanistic impact of data quality on machine learning performance for tabular data. Addressing four core data defects—missingness, noise, inconsistency, and redundancy—we propose a unified framework that decouples multidimensional quality issues and introduce an interpretable quality–performance sensitivity analysis. Leveraging both synthetic contamination and real-world degradation experiments, we evaluate 19 algorithms—including XGBoost, MLP, and TabTransformer—across classification, regression, and clustering tasks. Causal diagnosis is enabled via SHAP-based attribution and error溯源. Key findings: label noise and feature missingness exert the strongest negative effects; high-quality preprocessing improves average downstream F1 by 12.3%; and we establish, for the first time, quantitative relationships between data quality thresholds and model robustness—providing empirical foundations for data governance and algorithm selection.

Problem

• Examining impact of data quality on ML algorithm performance
• Analyzing six data quality dimensions across 19 ML algorithms
• Investigating pollution effects in training, test, or both datasets

Innovation

• Explores data quality impact on ML performance
• Tests 19 algorithms across classification, regression, clustering
• Analyzes three pollution scenarios in AI pipeline