CertainPrep: Showcasing Minimal Imputations in ML Data Preparation

Real-world data often contains missing values, outliers, and inconsistency. To train accurate models over real-world datasets, users need to spend a substantial amount of time and resources imputing and finding proper values for cleaning data items. In this paper, we demonstrate that it is possible to learn accurate models without data imputations for certain training data and target models. We propose a unified approach for checking the necessity of data imputation to learn accurate models across various widely-used machine learning paradigms. Our CertainPrep system, built upon this approach, checks this necessity efficiently with theoretical guarantees and returns accurate models in cases where imputation is unnecessary. When imputation is required, the system recommends minimal imputations. Our demonstration emphasizes interactivity and performance characterization to showcase the time and effort savings from data imputation while delivering accurate ML models.