This paper highlights that while machine learning model performance heavily relies on the quality of input data, real-world applications often face data-related challenges. Specifically, we address the common problem of distributional differences between two datasets collected from the same domain. While existing techniques for detecting distributional differences exist, a comprehensive approach that goes beyond opaque quantitative metrics and explains these differences in a human-readable way has been lacking. To address this, this paper proposes a multi-interpretable methodological framework for dataset comparison. Through various case studies, we demonstrate the effectiveness of this methodology across various data types and dimensions, including tabular data, text data, images, and time-series signals. This methodology complements existing techniques to provide actionable and interpretable insights that help understand and address distributional shifts.
