Classification of System Dynamics model outputs using decision trees

Classification of behaviours generated by mathematical models such as an ODE is an important component in modelling fields including System Dynamics. Useful for model validation, and investigation into the parameters which drive the different behaviours, a machine learning model based on a decision tree is a valuable way to interpret the behaviours returned. This research presents the creation of categorical attributes for the classification of model outputs into 13 behaviours. With a pre-given training set, it allows for the classification of unlabelled data, with hyper-parameters which can be changed for fine-tuning depending on the model presented. Where asymptotic model outputs may cause difficulty, a user-defined threshold value is available. Tested using empirical data, the results show a strong improvement on the previously available methods for behaviour classification of System Dynamics model outputs, and demonstrated using F1 scores. Our method has general applicability for classification of all time series data. • Classification of time-series data into 13 generic behaviours using a decision tree. • Pre-trained dataset allows for classification of unlabelled data of time-series data into behaviours. • Categorical attributes used with hyper-parameters improves on previous classification methods of time-series behaviours in system dynamics field.