Overview

Introduction to Predictive Process Monitoring

Business processes, ranging from the assembly line of a car manufacturing plant to the chaos of a hospital emergency department, govern the day-to-day operations of companies and institutions. The ability to understand and analyze the process flow and, most importantly, predict how it will evolve over time is critical to optimize resource utilization, minimize waiting times, reduce costs, improve services and products quality, and achieve many other benefits to both organizations and users.

Predictive process monitoring (PPM), a subfield of process mining, focuses on tackling all these predictive problems. For this purpose, PPM techniques train a predictive model using event logs, which provide information about the event execution of the business process, including the case identifier, the recorded activity, its timestamp, and various attributes, such as the resource information, as shown in Table 1. Although there are some datasets containing publicly available real-life event logs, each author tends to utilize a distinct subset of these under widely different experimental setups, which complicates the task of making a meaningful comparison between approaches.

Table 1 Extract from a real event log.

CaseID	Activity	Timestamp	Resource
Case3814	Assign seriousness	2012-05-04 15:23:08	Value 1
Case3814	Take in charge ticket	2012-05-07 07:01:12	Value 6
Case3814	Resolve ticket	2012-05-24 15:25:23	Value 12
Case3814	Closed	2012-06-08 15:25:48	Value 5
Case3815	Assign seriousness	2010-09-08 10:36:14	Value 2
Case3815	Wait	2010-09-08 15:32:25	Value 14
Case3815	Take in charge ticket	2010-09-13 09:01:56	Value 2

The PPM workflow

Fig. 1 Predictive monitoring workflow

Fig. 1 shows the usual workflow in PPM. Normally, predictive monitoring approaches receive as inputs a sequence of events from a running case up to a certain point (event prefix). This information serves as historical knowledge which is then used to train a predictive model with the objective of forecasting any variable of interest from the future events that will happen after that point in time (event suffix). Some of the most common predictive tasks include predicting the next event or sequence of events to the end of the case, the remaining time until the end of the case, or the future sequence of activities to be executed (activity suffix). Note that the fact of extracting the prefixes and their corresponding targets is a time-consuming task that often takes a large part of the approach coding effort and is highly prone to software bugs.

Benchmarking deep learning-based approaches

A plethora of machine learning techniques have been applied to the predictive monitoring problem, such as decision trees, support vector machines, and, more recently, deep learning-based architectures, with the latter being the most successful and popular. Despite the abundance of new approaches, most of them rely on very different experimental setups, so it is hard to draw definitive conclusions about their performance according to the inner characteristics of the event log. To tackle this issue, [1] presents an empirical comparison of the most relevant deep learning-based predictive monitoring approaches, standardizing data partitioning type and performance metrics, and including statistical tests to determine statistically significant differences between proposals. Unfortunately, this approach is difficult to replicate and requires a significant amount of work to integrate into the usual workflow of a data scientist or researcher who intends to develop a new predictive monitoring approach.

The importance of VERONA library

The VERONA library aims to address all of the mentioned challenges by implementing the necessary functions to develop the PPM workflow and reproduce the benchmark of [1] in an easy and user-friendly manner. This library provides primitives for event log retrieval and pre-processing, data partitioning, evaluation with benchmark metrics, and visualization of results with plots. Furthermore, it also includes the R implementation of the Bayesian statistical comparison proposed in [2] into the Python language, which is one of the core pillars of the benchmark. Finally, it allows importing the results of the original benchmark as well as three additional approaches, so that any new approach can be easily compared with the ones already evaluated. Designed for easy integration, VERONA facilitates the encoding, execution, and evaluation of common predictive monitoring tasks with minimal code.

References

[1] (1,2)

Efrén Rama-Maneiro, Juan C. Vidal, and Manuel Lama. Deep learning for predictive business process monitoring: review and benchmark. IEEE Transactions on Services Computing, 16(1):739–756, 2023. doi:10.1109/TSC.2021.3139807.

[2]

Borja Calvo and Guzmán Santafé. Scmamp: statistical comparison of multiple algorithms in multiple problems. R J., 8(1):248, 2016. URL: https://doi.org/10.32614/rj-2016-017, doi:10.32614/RJ-2016-017.