Time-series explainable artificial intelligence for real-world applications

Abstract

As the usage of deep learning models rapidly grows in real-world industries, the need for time-series explainable modeling increases in order to support human's complex decision-making process. In this thesis, I introduce practical problems that arise when applying explainable models to real-world industries and discuss how uncertainty modeling, active learning, counterfactual inference-based approaches tackle these challenges with three perspectives. First, to tackle the limited learning environment with quantity and quality of real-world data, I introduce uncertainty-aware network that provides reliable future prediction and explanations that considers the notion of uncertainty. Second, in order to improve model explainability which agrees more with real-world practitioners, I propose an interactive attention learning framework. Lastly, I extend the conventional explainable approach to scenario-based explainable framework which provides scenario-based forecasting explanations based on the condition of desired counterfactual questions, which enables to help complex decision-making process of real-world practitioners by providing actionable information.