Risk-based data validation in machine learning-based software systems

Abstract

Data validation is an essential requirement to ensure the reliability and quality of Machine Learning-based Software Systems. However, an exhaustive validation of all data fed to these systems (i.e. up to several thousand features) is practically unfeasible. In addition, there has been little discussion about methods that support software engineers of such systems in determining how thorough to validate each feature (i.e. data validation rigor). Therefore, this paper presents a conceptual data validation approach that prioritizes features based on their estimated risk of poor data quality. The risk of poor data quality is determined by the probability that a feature is of low data quality and the impact of this low (data) quality feature on the result of the machine learning model. Three criteria are presented to estimate the probability of low data quality (Data Source Quality, Data Smells, Data Pipeline Quality). To determine the impact of low (data) quality features, the importance of features according to the performance of the machine learning model (i.e. Feature Importance) is utilized. The presented approach provides decision support (i.e. data validation prioritization and rigor) for software engineers during the implementation of data validation techniques in the course of deploying a trained machine learning model and its software stack.