Architectural Overview

The replayer streams an event log in the order of timestamps (i.e. events are streamed in the order of their arrival, with the interarrival time being equal or proportional to that of the log). The arrived events are stored in the database and are sent to the message queue. The job worker retrieves the events belonging to the process instance in question (i.e a partial case) and sends them as an input to each pre-trained predictive model. The models output various metrics, such as the remaining time of an instance, its expected type of an outcome, next activity to be performed, etc. These metrics are visualized in the clients in the form of a dashboard table.

Framework architecture:

The following figure outlines the structure of a predictive engine. Firstly, we select a historical set of completed process cases. Then we encode them as feature vectors. These feature vectors are then fed to machine learning classifiers to train a model that is saved on a disk for later use. At runtime, given an ongoing process case, we encode it with the same encoding method as for the training, and apply the previously saved model. The engine returns a prediction result, e.g. the remaining execution time or a probability of future deviance.

Framework architecture: