ModelLens: An interactive system to support the model improvement practices of data science teams

Abstract

This demo presents ModelLens, an interactive system designed to support data science teams in their model improvement practices. A central component of improving models is analyzing model errors, ofen incorporating feedback on the model's precision and recall performance (e.g., diferences between the model's predicted label and its actual label). Today, error analysis is typically an ad hoc, team-specific process, largely accomplished through spreadsheets. ModelLens ofers a unified view of feedback from multiple sources, the ability for data scientists to explore the context of an individual feedback instance, as well as a customizable ontology to enable collaborative and systematic annotations of model errors.