Data Preprocessing Methods for Automating MLOps Pipelines: A Comparative Study

Preprocessing tools for data are increasingly being utilized in MLOps pipelines to develop models automatically. However, the fairness and reliability of automated processes are inadequately researched, risking causing performance degradation or bias. This discrepancy is addressed in this thesis with an evaluation of automated data preprocessing methods compared to a baseline approach, designed for integration into a TensorFlow Extended (TFX) pipeline. The performance of each method was compared in terms of classification measures and subgroup fairness to determine potential bias. Significance tests were employed to compare the performance of each automated method against the baseline. The results indicate that around half the automated methods had performance comparable to the baseline model, while the others performed much worse; more crucially, none of the automated methods significantly outperformed the baseline. These results show that not all preprocessing methods in automation can be used without manual validation.