Data Understanding for Data-Centric AI

Abstract Organizations collect growing volumes of data to extract value through analytics. However, this data growth creates challenges for effective data understanding, which forms the foundation for reliable decision-making and effective AI systems. Established analytics frameworks such as CRISP-DM and KDD acknowledge this importance but provide limited guidance to achieve this understanding, particularly for data-centric AI requiring collaboration across stakeholder groups. To address this gap, the authors conducted a systematic literature review, developing a five-dimensional framework for data understanding. They then performed a systematic mapping study analyzing how existing methods support these dimensions and accommodate different target audiences. The analysis reveals critical gaps in current methods, particularly in systematically supporting the understanding of data collection and contextualization. While most methods target data experts, the authors find a notable lack of methods supporting domain experts and decision-makers. This research advances both theoretical understanding by identifying the key dimensions that constitute data understanding and practical implementation by providing organizations with guidance on building data understanding.