Abstract for: Automated Discovery of Polarity from Data in System Dynamics Context

In any modeling field, one tries to use data as much as possible, since model construction time and model subjectivity can be reduced by data analysis. In this research, our focus is to perform automated data analysis to determine the polarity of the links between variables, under the assumption that we know all variables influencing a given effect variable. We propose an algorithm, discoverpolarity, which is tested with 7680 different synthetic data sets. Then the results are compared with Shape Constrained Additive Model and Pearson’s correlation analysis. Discoverpolarity clearly outperforms correlation analysis and slightly outperforms SCAM in terms of returning the correct polarity. If the underlying structure is multiplicative, it outperforms both benchmark methods in terms of returning the correct unique polarity. Our algorithm yields promising results when the dataset is rich and diverse. However, when the data consists of all perfectly correlated points, the algorithm tends to return multiple results, including the correct result and some wrong results. Another limitation is that the modeler must input the proper threshold values used in the algorithm.