The Impact of Data Quality on Artificial Intelligence Model Performance: An Empirical Study

The reliability of machine learning systems depends not only on the algorithms used but also on the quality of the data they learn from. In real-world applications, datasets often contain missing values, noisy features and incorrectly labelled instances. These problems can reduce model performance, sometimes more than expected. This study examines how common data quality issues affect predictive performance through a series of controlled experiments. We used three publicly available datasets: Breast Cancer Wisconsin, Wine Recognition, and the Diabetes dataset, retaining its original continuous target. Logistic Regression and Random Forest Classifier were tested on the first two datasets, while Random Forest Regressor was used for Diabetes. We introduced missing values (10% to 50%), feature noise (10% to 40%), and classification label noise (10% to 25%). For regression, we added target noise at the same three levels. Classification performance was measured using accuracy and macro-F1; regression performance was assessed using MAE, MSE, RMSE and R-squared. Each condition was repeated with ten different random seeds. The results show that the effects of data quality problems vary across datasets and tasks. In the classification experiments, label noise had the strongest effect. When 25% of the training labels were incorrect, average accuracy across the two models fell from 0.9655 to 0.9056 on Breast Cancer and from 0.9843 to 0.9111 on Wine Recognition. The changes in Diabetes regression were less pronounced. RMSE increased from 58.3157 on clean data to 59.4624 when 40% of the training values were missing. Feature noise produced slightly lower RMSE values in some cases, while target noise changed the score by less than 0.6. These differences were small compared with the variation across the ten runs and should therefore be interpreted with caution. We did not carry out formal statistical significance tests. Overall, the study examines two classification models and one regression model, and it also proposes a simple four-step approach to help identify data quality problems and understand their effects before investing more time in model development. The findings highlight the importance of data quality in machine learning workflows. Systematic assessment and basic preprocessing should be treated as essential steps, rather than optional tasks that can be overlooked in favour of more complex algorithms.

Authors

Institutions

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-15
DOI
https://doi.org/10.5281/zenodo.24243098
Primary Topic
Machine Learning and Data Classification
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

The Impact of Data Quality on Artificial Intelligence Model Performance: An Empirical Study

ANJALI GOSWAMI, ELIAS TAFADZWA MUDZUDZDU
Zenodo (CERN European Organization for Nuclear Research)
Machine Learning and Data Classification
preprint

The Impact of Data Quality on Artificial Intelligence Model Performance: An Empirical Study

ANJALI GOSWAMI, ELIAS TAFADZWA MUDZUDZDU
preprint en

Abstract

The reliability of machine learning systems depends not only on the algorithms used but also on the quality of the data they learn from. In real-world applications, datasets often contain missing values, noisy features and incorrectly labelled instances. These problems can reduce model performance, sometimes more than expected. This study examines how common data quality issues affect predictive performance through a series of controlled experiments. We used three publicly available datasets: Breast Cancer Wisconsin, Wine Recognition, and the Diabetes dataset, retaining its original continuous target. Logistic Regression and Random Forest Classifier were tested on the first two datasets, while Random Forest Regressor was used for Diabetes. We introduced missing values (10% to 50%), feature noise (10% to 40%), and classification label noise (10% to 25%). For regression, we added target noise at the same three levels. Classification performance was measured using accuracy and macro-F1; regression performance was assessed using MAE, MSE, RMSE and R-squared. Each condition was repeated with ten different random seeds. The results show that the effects of data quality problems vary across datasets and tasks. In the classification experiments, label noise had the strongest effect. When 25% of the training labels were incorrect, average accuracy across the two models fell from 0.9655 to 0.9056 on Breast Cancer and from 0.9843 to 0.9111 on Wine Recognition. The changes in Diabetes regression were less pronounced. RMSE increased from 58.3157 on clean data to 59.4624 when 40% of the training values were missing. Feature noise produced slightly lower RMSE values in some cases, while target noise changed the score by less than 0.6. These differences were small compared with the variation across the ten runs and should therefore be interpreted with caution. We did not carry out formal statistical significance tests. Overall, the study examines two classification models and one regression model, and it also proposes a simple four-step approach to help identify data quality problems and understand their effects before investing more time in model development. The findings highlight the importance of data quality in machine learning workflows. Systematic assessment and basic preprocessing should be treated as essential steps, rather than optional tasks that can be overlooked in favour of more complex algorithms.

Zenodo (CERN European Organization for Nuclear Research)
Kalinga University (IN)
Machine Learning and Data Classification
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon