A Data-Centric Approach to Missing Data Imputation: Addressing Noise, Adversarial, and Fairness Challenges

Real-world datasets frequently suffer from quality issues such as noise, class imbalance, and missing values, all of which can jeopardize machine learning models performance. Aligned with the Data-Centric AI paradigm, this work focuses on missing data, investigating its interaction with other data quality challenges rather than proposing new imputation methods. Specifically, the study examines how missing data behaves in the presence of noise, adversarial attacks, and fairness-related concerns. The findings reveal that such interactions notably influence both predictive performance and fairness outcomes. Notably, adversarial attacks not only degrade model accuracy but also disrupt imputation effectiveness. These insights emphasize the importance of considering broader data quality factors when addressing missing data. Additionally, this research contributes to a novel Python package designed to generate missing values under various realistic settings. This tool facilitates reproducible experiments and enables more equitable benchmarking of imputation strategies, supporting future research in missing data and data-centric evaluations.