A system for storage and analysis of machine learning operations

Abstract

Data scientists go through an iterative process when building machine learning models. This process includes operations like feature selection, cross validation, model fitting, and evaluation, which are repeated until a sufficiently accurate model is produced. This thesis describes ModelDB Server and the ModelDB Spark Client, which record operations as the data scientist performs them, stores the operations and models in a central database, and exposes an API for gleaning information from the operations and models. ModelDB Server is library agnostic and it can serve as a foundation for other applications. ModelDB Spark Client is a library for the Apache Spark ML machine learning library that lets the data scientist log their operations and models with minimal code changes. ModelDB Server and Spark Client have low time and space overhead for large training dataset sizes and the overhead is independent of the dataset size.