XAI meets Data-Centric AI: The Need to Explain Data Processing

Data preprocessing is an essential part of any AI system, as indicated by the current trend of data-centric AI. However, until now, explainability efforts have almost exclusively focused on models. Choices in preprocessing, such as handling missing data, outliers, scaling, and class balancing, and their effects are currently not part of the explanations. If we continue this way, we are left with only a partial understanding of the model’s outputs. This transparency deficit creates risks for accountability, fairness, model performance, and more. We argue why we need full pipeline transparency and, as part of that, data preprocessing explainability, even in an age of AutoML and foundational models.