Activity-based modeling strategy for reliable machine learning system analysis targeting GUI-based applications

To address the requirement of analytical methods to ensure machine learning (ML) system reliability, a multi-view modeling framework for ML systems (M3S)[1] is proposed. This goal-oriented risk analysis framework uses metamodels to achieve consistency among various analytical models and can be extended to new models when necessary. To analyze the risk avoidance by human oversight, activity diagrams were combined with the model extensions of M3S [3]; however, the reasons for using activity diagrams, their characteristics, and their specific usage remain unclear. In this study, we examined the effectiveness of the M3S in systems involving collaborative work between ML and humans by defining a reference case of an optical character recognition (OCR) application as an example of a graphical user interface (GUI) application and describe the procedure to use activity diagrams in risk analysis through a case study. Specifically, we present a procedure for creating multiple models, including activity diagrams, while maintaining consistency with the other models. Furthermore, we demonstrated the evaluation of the changes in risk assessment values using activity diagrams by introducing risk assessment formulas into the activity diagrams and thus confirmed the effective analysis of risk improvement measures (countermeasures) based on human intervention.