# The Effects of Data Quality on ML-Model Performance

**Authors:** Lukas Budach, Moritz Feuerpfeil, Nina Ihde, Andrea Nathansen, Nele Noack, Hendrik Patzlaff, Hazar Harmouch, Felix Naumann  
**arXiv:** [2207.14529](https://arxiv.org/abs/2207.14529)  
**Date:** 2022-07-29  
**Summary page:** https://www.summarizepaper.com/en/paper/2207.14529v1/

## Key points

- The paper presents a comprehensive experimental study on the correlation between six data quality dimensions and the performance of fifteen machine learning algorithms.
- The need for high-quality training and test data is crucial for reliable AI applications.
- This paper is the first systematic study of the effects of data quality dimensions not only for classification but also for clustering and regression tasks.
- The authors perform a targeted analysis for cases where serving data, training data or both are of low quality.
- The results show that there is a strong correlation between all six traditional dimensions of data quality and ML-models performance across all three types of tasks (classification, regression, and clustering).
- All polluters used in their experiments are available online alongside all datasets used in their study which can be easily extended with further quality dimensions or ML models.

## Summary

This paper presents a comprehensive experimental study that investigates the correlation between six data quality dimensions and the performance of fifteen machine learning (ML) algorithms. The need for large quantities of training and test data in modern AI applications creates critical challenges not only concerning the availability of such data, but also regarding its quality. Incomplete, erroneous or inappropriate training data can lead to unreliable models that produce ultimately poor decisions. Trustworthy AI applications require high-quality training and test data along many dimensions, such as accuracy, completeness, consistency, and uniformity. While previous research has studied the effects of label noise and missing values on classification tasks, this paper is the first systematic study of the effects of data quality dimensions not only for classification but also for clustering and regression tasks. The authors perform a targeted analysis for cases where serving data, training data or both are of low quality. The authors present a systematic empirical benchmarking to understand the correlation between data quality and ML-models performance under the umbrella of data-centric AI. They simulate real-life scenarios concerning data in ML-pipelines and provide practical insights and learned lessons for data scientists. Additionally, they raise several questions and point out possible directions for further research. The paper distinguishes three scenarios based on the AI pipeline steps that were fed with polluted data: polluted training data, test data or both. The results of their experiments show that there is a strong correlation between all six traditional dimensions of data quality (accuracy, completeness, consistency, uniformity, timeliness and believability) and ML-models performance across all three types of tasks (classification, regression and clustering). Overall, this work is a first step towards linking ML-models performance to underlying data quality while also understanding their connection. All polluters used in their experiments are available online alongside all datasets used in their study which can be easily extended with further quality dimensions or ML models. This paper provides valuable insights into how to improve the reliability and trustworthiness of AI applications by ensuring high-quality training and test data.