The work compares the use of recurrent networks and models based on transformer architecture to solve the problem of predicting the completion time of a business process. Models, by definition, model the sequence of actions and are able to take into account a variety of attributes in determining target characteristics. For comparison, a recurrent model of long-term short-term memory and a transformer encoder of its own architecture were used, the operation of which was tested on openly presented real data from the logs of the support service. The models were trained and tested using Python using the pandas, numpy, and torch libraries with the same data preparation, prefix generation, and time division for both models. The comparison as a result of experiments on the average absolute error showed the advantage of the transformer encoder; approximately the same accuracy of the models with a slightly higher accuracy of the transformer model was recorded for the standard deviation.

Keywords: predictive monitoring, event log, machine learning, transformer encoder, neural networks, data preparation, regression model, normalization, padding, recurrent network, model architecture