Why did the prediction change? Explaining changes in predictions as time progresses

Abstract

Few works on machine learning (ML) explanations design explanations from the perspective of model deployment in the real-world. This work addresses the challenges of understanding ML models applied to event-based time-series data, concretizes two explanation scenarios, and proposes explanations based on changes in feature values, model predictions, and feature contributions for each deployment scenario. We study the prediction problem of turbine brake pad failures, where predictive time-series ML models were deployed in production. Our solution to help decision makers understand how the predictions are made include the development of a usable ML interface and explanations that are aware of the scenarios and contexts where the models are being used. We discuss the usage of ML explanations and the importance of the context under which the model is deployed. We showed our usable ML interface and the explanations with their corresponding scenarios built on top of the usable ML system, which consists of Pyreal, Sibyl-API, and Sibylapp.