As organizations gain awareness of the potential business value locked in their process execution event logs, \evidence-based" business process management (BPM) becomes a common tool for process analysts. In contrast to traditional process monitoring techniques which are typically performed using data from running process instances only, predictive evidence-based BPM methods tap also into historical data, to allow process workers to respond, in real-time, to speci c process performance issues and compliance violations as they arise or even before they arise. In previous work, various approaches have been proposed to address typical predictive process monitoring problems, such as whether a running process instance will meet its performance targets, or when will an instance be nally nished. However, these approaches are rather ad-hoc and lack generality, as they tackle only particular, pre-de ned aspects of predictive monitoring and often only work with speci c characteristics of the dataset. The proposed research project aims at developing a general and robust framework for predictive process monitoring that will address a variety of process monitoring tasks such as predicting the outcome of individual activities or of the whole process instance, or predicting the completion path of an instance.
Business process monitoring and controlling includes activities wherein data related to the process execution are collected and analyzed to assess the process performance with respect to its performance criteria.
Traditional approaches to process monitoring and controlling rely on \postmortem" (o ine) analysis of process execution or runtime observations of process work by process analysts. The former includes a family of techniques known as process intelligence, which takes as input a database of completed process cases and outputs process performance insights, such as identi ed bottlenecks or historical cycle times. In comparison, business activity monitoring is performed at runtime and takes as input an event stream, i.e. pre xes of running process cases. The output of the business activity monitoring is a real-time picture of the process performance, such as current process load or cases running late.
These techniques, while being able to provide estimations of the process
performance, are reactive in nature, meaning that they detect process issues
only once they have occurred. Predictive process monitoring aims to address
the limitations of traditional \empiric" monitoring practices by systematically
utilizing data produced during the process execution to continuously monitor the
processes performance [
Business activity monitoring
Process Monitoring and
Controlling online Event stream Predictive process monitoring
offline Event logs
Process intelligence
At the core of every predictive process monitoring system are prediction
models that are built for speci c prediction goals based on the event log L of
completed cases (Figure 2). These models are applied to pre xes p of running
cases in order to make predictions about their future performance. If the
predicted outcome deviates from the expected (normal) process behavior, alerts are
issued to notify the process participants. Additionally, recommendations can be
issued to the process workers, advising them of the impact of a speci c action
on the probability that the current case will violate the performance objectives.
Therefore, problems are anticipated and can be proactively managed. For
example, in a freight transportation process a prediction goal can be the occurrence
of delay in delivery time. In this case, the outcome will be whether the delay is
going to occur (or probability of delay occurrence). If a delay is predicted, faster
means of transport or alternative transport routes could be scheduled proactively
and before the delay actually occurs [
The rest of the paper is organized as follows. In Section 2, we summarize state of the art of predictive process monitoring. In Section 3 we indicate a research gap and formulate the problem to be solved, along with the research questions to be addressed and research criteria for the evaluation. Section 4 provides an overview of the proposed research plan. Next, Section 5 summarizes the progress so far and features a discussion on the preliminary results. Finally, Section 6 concludes the paper and discusses future work.
Prediction outcome(s) Event logs L Ongoing case p
Prediction goal(s)
Prediction model(s)
Predictive process monitoring has recently received signi cant interest, due to
the widespread adoption of work ow management systems with event logging
capabilities, as well as due to advances in process mining techniques. In this
section we classify the proposed techniques based on the prediction goal they
address, i.e. time-related predictions, predictions of an outcome of a process
instance and predictions of the process instance continuation or characteristics
thereof.
The rst group of works dealing with time-related predictions approaches them
as a sequence classi cation problem, focusing on deadline violations and delayed
process executions. For example, Pika et al. [
Other research proposals have approached time-related predictions as a
regression problem to predict the remaining processing time and cycle time of a
process instance. For instance, van der Aalst et al. [
Here one group of works focuses on predicting the occurrence of various kinds of
process faults. For instance, Kang et al. [
An outcome of a process case can also be de ned with respect to compliance
with prede ned rules (e.g. \every purchase order should eventually be followed
by an invoice issue") or service level agreements (e.g. \the amount of customer
complaints should not exceed 10% of customers"). In this regard, a paper by
Maggi et al. [
Most of the above-mentioned approaches focus on intra-case predictive
monitoring, meaning that running cases are viewed independently from each other.
However, in real-life scenarios, one needs to take into account dependencies
between cases, particularly due to resource contention and data sharing.
Additionally, the distribution of types of cases (e.g., high number of \di cult" open
cases) a ects the percentage of cases that end up in a negative outcome [
A range of research proposals has tackled the problem of predicting the
continuation of running process instances or characteristics thereof. For example,
Lakshmanan et al. [
De Leoni et al. [
As demonstrated before, a number of techniques have been proposed to address very speci c prediction goals, e.g. prediction of deadline violations or are applicable for speci c datasets. The working hypothesis of the project is that most prediction goals can be viewed as speci c instances of a more generic problem, that is prediction of a completion path of a given case (starting from a current point in time) and characteristics of that path. For example, to predict the occurrence of defect waste induced when a defective execution of an activity leads to a rework (i.e. a part of the process having to be repeated in order to correct the defect), one can query the obtained likely sequence completions for the presence of rework loops.
Consequently, in this project we aim to develop a general predictive monitoring framework that can be instantiated for a wide range of prediction goals and for di erent domains, e.g. transportation, insurance and healthcare domains. In other words, the goal of the project is to move the state of the art of predictive monitoring from ad-hoc solutions with a narrow focus to a general solution framework.
The proposed framework will take as input: (i) an event log of historical, i.e. completed process execution traces; (ii) a sequence of activities executed in a given process case along with its data attributes and (iii) a set of prediction goals (Figure 2). The output will be determined by the prediction goal. For example, if the goal is the prediction of the remaining processing time, the output will be the most likely value of the remaining time along with its con dence interval. If the goal is to predict the discrete outcome of a case, the framework will output the probabilistic distribution of the possible outcomes.
The research project will address the following research questions: { RQ1. How can predictive monitoring methods be used to predict the most likely outcome of the process? The outcome can be binary (e.g. whether the case ended up normally) or a numerical function (e.g. the cycle time of the case). { RQ2. How can predictive monitoring methods be used to predict the outcome of a speci c activity? Similarly to the previous problem, the outcome of an activity can be boolean (e.g. whether a check has been passed) or numerical (e.g. the cycle time of the activity). { RQ3. How can predictive monitoring methods be used to predict the entire sequence of activities leading to the process end and characteristics of the sequence? In other words, for an ongoing process case, we aim to determine the most likely trace su x.
The framework designed as a result of this research will be evaluated using the several criteria. First, the solution should be able to make predictions about a given outcome with relatively high levels of accuracy. Next, it should be possible to make predictions as early as possible, as soon as there is enough information in an uncompleted trace to make a prediction with su cient con dence. Finally, the predictions should incur a minimum runtime overhead to be made almost instantaneously, with limited computing resources, over real-scale scenarios.
Predictive process monitoring is usually used for the early classi cation of process instances into normal and deviant based on some performance metrics. However, explaining the reasons for a process case deviation from its expected execution lays foundation for deviance mining techniques, which are largely out of scope of the current PhD project.
Another limitation stems from the fact that prediction accuracy depends on the variability of the process. Speci cally, processes that signi cantly variate between cases are usually highly unpredictable. Hence, a reasonable prediction accuracy can be only achieved at the expense of earliness, i.e. only when the case is about to nish.
In addition, we will largely focus on intra-case predictive monitoring, meaning that we see a running case in isolation from the others, thus discarding inter-case relations, such as those due to resource and data contention. 4
The research approach of this project is inspired by the design science methodology. In line with the design science methodology (Figure 3), we started from a preliminary literature review of the predictive process monitoring approaches. This study allowed us to identify a gap in the current state of the art and formulate a research proposal to address the gap.
RQ1 RQ2 RQ3 1. Define a problem • Literature review • Problem formulation • Literature review 2. Suggest a solution • Pilot design
• Baselines & hypotheses 3. Develop artifacts 4. Evaluate artifacts 5. Conclude • Develop algorithms • Implement software • Conduct experiments to
validate the artifacts • Publish the results and
findings
The next stage involved further studies of related work with a particular focus on machine learning and process mining algorithms behind them. This stage resulted in construction of baseline predictive models and formulation of hypotheses to be checked at the later stages.
At the third stage of the project, we produce viable artifacts in the form of software that implements the models for predicting various properties of running business processes, e.g. whether the process case will meet a certain performance objective and whether the case will contain a re-work loop. Aside from predictions, the models will potentially be able to explain the causes of various case properties.
A research contribution requires a rigorous evaluation of the artifacts. Thus, during the fourth stage, we plan to validate the proposed techniques with respect to the previously de ned research criteria, using datasets exhibiting di erent characteristics. Speci cally, a trade-o between earliness and accuracy for various prediction goals and domains needs to be assessed. To test the applicability of the framework at runtime, we also have to measure how fast the predictions can be calculated. Additionally, we plan to compare predictive methods with reactive ones.
Finally, at the last stage of the project, we will wrap up the results of our studies in the form of a doctoral thesis (by monograph).
Stages 2, 3 and 4 constitute the core part of this research and will be implemented according to an iterative and incremental development model, meaning that they will be sequentially revisited for each research question. In addition, we plan to address research questions RQ1 and RQ2 in two major iterations. Initially, we develop a simpler technique to predict the case completion path, while during the next iteration we re ne the previously obtained results by lifting the assumptions made in the rst iteration. Thus, these stages will be partly overlapping. 5
In this section, we describe our present academic contribution, in line with the formulated research questions (see Section 3).
First, to address RQ1, we have developed a framework to predict the most
likely outcome of an ongoing case, given its pre x and a set of traces of
historical (completed) cases. Speci cally, we extended the approach presented
in [
Next, the research question RQ2 has been tackled in the context of process
optimization by minimizing overprocessing, i.e. the e ort that is spent in the
performance of activities to an extent that does not add value to the customer nor
to the business. Speci cally, we have developed a framework to predict, at
runtime, processing times and rejection probabilities of the checks to be performed
during the execution of a case. Hence, the checks can be resequenced at runtime
based on the characteristics of the current case in such an order that incurs the
lowest overprocessing. The evaluation of the proposed framework showed that
for processes with the reject probabilities being close to each other, it
outperforms a traditional design-time ordering approach, wherein checks are performed
in a xed order across all cases. The outcome of the stage has been formalized in
a paper titled \Minimizing overprocessing waste in business processes via
predictive activity ordering". The paper has been accepted for publication at the
CAiSE 2016 [
Finally, to address RQ3, we plan to develop a framework for predicting sequence completion for an ongoing process case. Speci cally, for a given point during the process case execution, we extract all possible scenarios of how this case may unfold up to its completion and estimate probabilities of each scenario. This stage of the project is currently in the implementation phase. We utilize a two-module framework wherein the rst module, for a current ongoing execution trace, nds equivalent or similar pre xes in the historical log of completed traces; while the second module analyzes su xes corresponding to previously discovered pre xes and learns a structured prediction model using techniques like recurrent neural networks or hidden Markov models. First experiments have shown that, for most logs, predictions are accurate and robust only with short su xes, i.e. only for the next few events. 6
In this paper, we detailed and motivated our PhD project. A research gap in state of the art predictive business process monitoring has been identi ed and the outline of the research plan for investigating a solution for the identi ed gap is proposed. In future work, the so far developed artifacts, along with the technique for the prediction of case completion path currently being developed { will be incorporated into a general framework for predictive process monitoring that would be able to address speci c process monitoring tasks. The framework will be implemented using a combination of state-of-the-art machine learning and process mining algorithms, and evaluated using real-life datasets with di erent characteristics.