Privacy preserving explanations for hierarchical time series forecasts

Abstract

Data privacy and explainability are two important requirements for any mature AI enabled system. Local explainability for a prediction or forecast amounts to assigning credit or blame to different input features of a model responsible for that prediction. Aggregation of these predictions and explanations to higher levels of hierarchy is often met with the challenge of privacy loss as it reveals characteristics of individual data points to a wider audience. Hence, an optimal tradeoff between privacy and explainability is explored in the context of hierarchical time series forecasting.