Efficient Computation of ML Explanations using Provenance

Understanding machine learning (ML) models is essential as it fosters trust, transparency, and the ability to identify and address any issues that may arise.This paper proposes a method for generating widely recognized ML explanations based on provenance.We note that commonly used ML models can be represented as a set of Datalog queries.These queries can involve complex operations, including recursion, reflecting the iterative nature of training ML models.We explore the process of rewriting these queries to capture provenance and to compute explanations efficiently during the model training process.