PROXAI: Interactive Provenance-Aware Debugging of Machine Learning Pipelines
TL;DR
PROXAI, an interactive system that uses XAI techniques to identify influential data elements and follows their lineage through a multi-granular provenance graph, enabling users to trace preprocessing root causes, inspect transformation logic, and validate corrective actions within a unified debugging workflow.
Abstract
Data provenance is a well-established foundation for transparency and auditability in data management, yet existing provenance systems offer little support for connecting lineage information to downstream model behavior. Conversely, Explainable AI methods identify influential features and training instances, but treat the training data as a fixed input, leaving their preprocessing history unexamined. We present PROXAI , an interactive system that bridges these two perspectives to support provenance-aware debugging of machine learning pipelines. Starting from anomalous model behavior, PROXAI uses XAI techniques to identify influential data elements and follows their lineage through a multi-granular provenance graph, enabling users to trace preprocessing root causes, inspect transformation logic, and validate corrective actions within a unified debugging workflow. Our demonstration highlights how combining XAI and provenance enables end-to-end explanation and practical debugging of data-centric AI pipelines.