Publikation
Demonstrating PIPE-X: Supporting Iterative Pipeline Development Through Explanations
Nadja Geisler; Benjamin Hättasch; Carsten Binnig
In: Wolfgang Lehner; Vanessa Braganholo; Kostas Stefanidis; Zheying Zhang; Alexander Krause; Jo~ao Felipe Nicolaci Pimentel (Hrsg.). Proceedings 29th International Conference on Extending Database Technology, EDBT 2026, Tampere, Finland, March 24-27, 2026. International Conference on Extending Database Technology (EDBT-2026), March 24-27, Tampere, Finland, Pages 720-723, OpenProceedings.org, 2026.
Zusammenfassung
Building good data processing systems from an end-to-end perspective is as important as it is complex. While explainable AI (XAI) approaches help users to understand the behavior of trained models, choices in preprocessing and their effects are not yet part of explanations. We have recently proposed PIPE-X (Preprocessing Impact and Pipeline Explanations) to explain data preprocessing pipelines by calculating the impact of each step on the model’s behavior. In this paper, we demonstrate how data scientists can leverage PIPE-X through an interactive graphical interface to gain insights into their pipeline and improve the end-to-end system accordingly. Users can provide their data, pipeline, and model to obtain the impacts for individual model outputs or an overview of the effect throughout the model. They can leverage various interactive outputs (graphical and numerical) to gain the best result for their specific use case.