Interpreting Intelligibility under Uncertain Data Imputation

Many methods have been proposed to make machine learning more interpretable, but these have mainly been evaluated with simple use cases and well-curated datasets. In contrast, real-world data presents issues that can compromise the proper interpretation of explanations by end users. In this work, we investigate the impact of missing data and imputation on how users would understand, and use explanation features and propose two approaches to provide explanation interfaces for explaining feature attribution with uncertainty due to missing data imputation. This work aims to improve the understanding and trust of intelligible healthcare analytics in clinical end users to help drive the adoption of AI.