Explaining a Random Forest With the Difference of Two ARIMA Models in an Industrial Fault Detection Scenario
|A. Glock. Explaining a Random Forest With the Difference of Two ARIMA Models in an Industrial Fault Detection Scenario. volume 180, pages 476-481, DOI https://doi.org/10.1016/j.procs.2021.01.360, 2, 2021.|
|Buch||Procedia Computer Science|
In this paper a method is proposed to obtain explainability of the random forest model. Two Auto-Regressive Integrated Moving Average (ARIMA) models form the basis for this approach. The ARIMA models are used in a way similar to how local surrogate models are typically applied. The explanation of random forest’s prediction is derived from the numerical differences of the parameters of the ARIMA models. To demonstrate the feasibility of this idea, an experiment that implements this approach is conducted. The data used for this are similar to an accumulated bathtub curve representing failure rates in a production process. The results of the experiment show that the approach is able to identify a linear trend in some parts of the data, and therefore locally provide an explanation for the functional form of the underlying failure rate.