ML Lineage for trustworthy machine learning systems

Abstract

Machine learning (ML) has reached technological maturity, and its applications are now becoming pervasive across diverse industries. Simultaneously, demands from society and authorities have become increasingly complex and stringent, requiring transparency and evidence for trustworthiness. While emerging MLOps practices streamline the development and operations of ML-based systems, the required information remains disconnected and insufficiently addresses diverse concerns. We contribute the concept of ML lineage as a framework to holistically capture and connect the required information about ML model development and operations. ML lineage fundamentally distinguishes between the model and prediction levels, conceptually encompassing separate yet interconnected core domains for the project, experiment, model, and prediction.