Elevating Business Process Modeling with QuerySelector: Use case with the BPI’12 dataset

Many software systems designed for Business Processes (such as Enterprise Resource Planning (ERP) systems for instance) are recording so-called event logs. Those events can be aggregated into whole cases i.e. (from the beginning of a Business Process to the end of it) as a sequence of events relevant to this process case.

Predictive business process methods utilize those sequences of event logs to make predictions about various future events of given cases such as:

predicting the next activity,
predicting the future path (continuation) of a running case,
predicting the remaining cycle time, predicting deadline violations,
predicting the fulfillment of a property upon completion.

Traditionally, those predictions were obtained by using simple statistics methods tailored to a particular business process. In recent years those problems attracted the attention of Deep Learning research and it turned out that they yield excellent results in many practical scenarios.

One of those examples is an analysis of a publicly available data set called BPI’12. This dataset was acquired from „The 2012 Business Processing Intelligence Challenge” (BPI 2012).

It contains real-life event logs of process events of loan or overdraft applications from a bank in the Netherlands over a period of more than 6 months. BPI’12 dataset is large enough to allow for applying Deep Learning methods – it contains a total of 262,200 events within 13,087 cases, starting with a customer submitting an application and ending with the eventual conclusion of that application into an Approval, Cancellation, or Rejection events. However, the sequence of events containing a single business case can be quite complicated and contain events of a very different nature ie. events related to the interaction of the bank with a client or bank’s inner dealings with consumer application. Additionally, the IT system controlling and monitoring this business process generates its own logs related to the overall stage of a given business case so events are put in the context of three transition modes – SCHEDULE, START, and COMPLETE. Since each event has its own unique timestamp we can order any such case on the timeline. Surprisingly, the above-described setup can be encoded into vector embeddings, and subsequently, Deep Learning methods can be applied to this problem.

In the case of the BPI’12 dataset, the first Deep Learning algorithm applied was the so-called Long Short-Term Memory (LSTM) neural network.

The prediction tasks were the following:

Prediction of the next type of activity to be executed in a running process instance
Prediction of the timestamp of the next type of activity to be executed
Prediction of the continuation of a running instance
Prediction of the remaining cycle time of an instance

It achieved very good results on all of those tasks but our own version of the Transformer algorithm – Query Selector – achieved even better (SOTA) results on the same tasks. Namely, it reaches an Accuracy of 0,79 on some tasks which shows that commercial applications of Deep Learning algorithms such as Query Selector have enormous business potential. It would lead to substantive cost-cutting and optimization of even very complicated business processes.