Machine Learning (ML) has become essential in today's production environments, which aim for high-performance models. However, the performance of these models highly depends on the data quality of the underlying data. Therefore, data preparation pipelines are used to ensure high data quality standards. In the process of ensuring data quality, it is possible to unintentionally introduce unfairness that favors a specific group. This can lead to serious implications for decision-making processes.

In this work, we address the challenge of ensuring fairness in data preparation pipelines by developing a system that alerts users if potential unfairness may have been introduced in one or multiple steps of the pipeline. Through a comprehensive literature review, we identify fairness metrics commonly used and investigate the relationship between fairness and typical data characteristics. Based on this understanding, we develop a system that aims to detect steps in which unfairness may have been introduced within the data preparation pipeline if the resulting ML model cannot satisfy certain fairness metrics.

In terms of fairness, we evaluate the performance of our proposed system across diverse benchmark datasets and various data preparation pipelines. We find that our system is a suitable method to detect steps within a data preparation pipeline that have potentially negatively influenced the fairness of the resulting model and thereby provide valuable feedback to the users. However, we also acknowledge the limitations of our work, including the focus on binary classification tasks and the consideration of only selected fairness metrics and data characteristics.

In conclusion, our work contributes to the field of data preparation pipelines by offering a framework designed to ensure fairness within data preparation pipelines. By addressing the intersection of fairness and data preprocessing, we aim to foster fairer decision-making processes for ML models in real-world applications.