Database and ML-driven software tool to predict CNC for BPMN models using TNN and TNSF

Abstract

The relevance of the study is determined by the growing role of business process modeling using BPMN for analyzing, improving, and automating the activities in organizations, where excessive structural complexity of models negatively affects their comprehensibility and practical value for users. The aim of the work is to develop and experimentally verify an objective approach to assessing the quality of BPMN models from the perspective of structural complexity based on formal graph properties, quantitative metrics, and machine learning methods. In the study, the business process is presented as a directed labeled graph, for which basic size metrics and the CNC network connectivity coefficient are used as key complexity characteristics. A domain specific model for CNC prediction based on linear regression and a mechanism for classifying BPMN models by complexity levels using threshold values are proposed. Experimental verification was performed on 6137 business process descriptions from the Camunda open repository and showed adequate prediction accuracy with MSE of 0.02 and R-squared of 0.71. The classification results demonstrated high efficiency for BPMN models of moderate and high complexity with an F1-score of up to 0.98. It was concluded that the combination of structural metrics and machine learning provides an objective and scalable assessment of the quality of BPMN models in terms of their understandability to users, therefore, business process models quality.