Outcome-Preserving Input Reduction for Scientific Data Analysis Workflows

Abstract

Analysis of data is the foundation of multiple scientific disciplines, manifesting in complex and diverse scientific data analysis workflows often involving exploratory analyses. Such analyses represent a particular case for traditional data engineering workflows, as results may be hard to interpret and judge whether they are correct or not, and where experimentation is a central theme. Oftentimes, there are certain aspects of a result which are suspicious and which should be further investigated to increase the trustworthiness of the workflow’s outcome. To this end, we advocate a semi-automated approach to reducing a workflow’s input data while preserving a specified outcome of interest, facilitating irregularity localization by narrowing down the search space for spotting corrupted input data or wrong assumptions made about it. We outline our vision on building engineering support for outcome-preserving input reduction within data analysis workflows, and report on preliminary results obtained from applying an early research prototype on a computational notebook taken from an online community of data scientists and machine learning practitioners.