Explainability versus privacy for hierarchical time series forecasting

Abstract

Sales and inventory data in supply chains has spatio-temporal characteristics, leading to a hierarchical time series framework. Accurate forecasting and decision making in a hierarchical framework depends on the data shared amongst participants from the same hierarchical level and also data sharing across different levels of hierarchy. Although conceptually attractive, many participants in the supply chain are unwilling to share detailed information due to the fear of unfair exploitation by other competing parties. There would also be regulatory requirements that limit the data sharing outside the data owner in plain form. Hence, privacy preserving protocols are the need of the hour. Apart from privacy, explainability is another essential requirement imposed on the design of most AI based systems. Supply chain participants who rely on hierarchical forecasts, expect to understand why are the forecasts high or low? Which adjacent node or a faraway node in the hierarchy is responsible for a change in the forecast? These requirements are in direct opposition to the privacy requirements of the participants, thus posing several engineering and research challenges. In this work, we focus on developing forecasting systems that exploit hierarchical data structures while optimizing for a trade-off between privacy and explainability.