Prescriptive process monitoring: Quo vadis?

Introduction

Process mining is a family of techniques that facilitate the discovery and analysis of business processes based on execution data. Process mining techniques use event logs extracted from enterprise information systems to, for instance, discover process models (Agostinelli et al., 2019) or to check the conformance of a process with respect to a reference model (Van Der Aalst, 2012). In this setting, an event log is a dataset capturing the step-by-step execution of a business process and includes timestamps, activity labels, case identifiers, resources, and other contextual attributes related to each case or each step within a case.

Over time, the scope of process mining has extended to encompass other use cases (Milani et al., 2022) such as techniques for predicting the outcome of ongoing cases of a process based on machine learning models constructed from event logs (Maggi et al., 2014; di Francescomarino et al., 2018). Predictions, however, only become useful to users when they are combined with recommendations (Márquez-Chamorro, Resinas & Ruiz-Cortéz, 2018). In this setting, prescriptive process monitoring is a family of methods that recommends interventions during the execution of a case that, if followed, optimize the process with respect to an objective (Shoush & Dumas, 2021). For instance, an intervention might improve the probability of the desired outcome (e.g., on-time delivery) or mitigate negative outcomes (e.g., delivery delays) (Metzger, Kley & Palm, 2020). Different implementations of prescriptive process monitoring have been proposed in the literature. However, the understanding of prescriptiveness in the field varies. To this end, some methods are guiding the user during the execution of the case based on its similarity to previous executions (Terragni & Hassani, 2019). Other methods specifically aim at optimizing process performance through correlation-based (Gröger, Schwarz & Mitschang, 2014; Ghattas, Soffer & Peleg, 2014) or causality-based predictions (Bozorgi et al., 2021; Shoush & Dumas, 2021). The methods also differ in other aspects, such as interventions prescribed. In some cases, two different methods aim at achieving the same objective but in different ways. For instance, to avoid an undesired outcome, one method might prescribe assigning resources for the next task (Sindhgatta, Ghose & Dam, 2016), whereas another might recommend which task to execute next (de Leoni, Dees & Reulink, 2020).

The benefits of prescriptive process monitoring can only be fully realized if these methods prescribe effective interventions that are followed (Dees et al., 2019). At present, though, the variety and fragmentation of prescriptive monitoring methods makes it difficult to understand which method is likely to be most effective or more likely to be accepted by end-users in a given business situation. There is no overview that captures existing prescriptive monitoring methods, what objectives they pursue, which interventions they prescribe, which data they require, or the extent to which these methods have been validated in real-life settings. Research overviews and classification frameworks have been put forward in the related field of predictive monitoring (di Francescomarino et al., 2018; Márquez-Chamorro, Resinas & Ruiz-Cortéz, 2018) and automated resource allocation (Arias et al., 2018; Pufahl et al., 2021). However, such works do not provide a structured overview of existing prescriptive process monitoring methods, nor serve researchers with a base for uncovering underserved but potentially valid areas of research in this field.

To address this gap, we study five research questions:

• Given that prescriptive process monitoring methods aim at prescribing interventions that produce business value, i.e., achieve an objective, we formulate the first research question as:   RQ₁. What is the objective for using prescriptive process monitoring methods to optimize a process?

• The second research question aims at discovering how these objectives can be achieved:   RQ₂. What are interventions that are prescribed by prescriptive process monitoring methods?

• Third, we explore the data required by the proposed methods:   RQ₃. What data do prescriptive process monitoring methods require?

• The fourth research question explores what modeling techniques the methods utilize to make use of the input data:   RQ₄. What modeling techniques do prescriptive process monitoring methods use?

• Finally, we explore under which policy interventions are prescribed:   RQ₅. What policies do prescriptive process monitoring methods use?

To answer these questions, we conducted a systematic literature review (SLR) following the guidelines proposed by Kitchenham & Charters (2007). We identified 37 papers that we analyzed to develop a multi-dimensional framework to characterize prescriptive monitoring methods. The contribution of this article is threefold. First, we provide a review of existing prescriptive process monitoring methods. Second, we develop a framework that classifies prescriptive process monitoring methods according to their objective, metric, intervention types, techniques, data inputs, and policies to trigger interventions. Third, we outline existing research gaps and provide insights into potential areas for future research in the field of prescriptive process monitoring. Our contribution aims at supporting researchers of prescriptive process monitoring methods. Researchers benefit from this contribution as they can gain insight into the current state of the art of the field and identify potential directions for future research. Developers of process mining tools who are interested in incorporating prescriptive process monitoring into their tools can also benefit from this work by better understanding the limitations and perspectives of existing methods.

The remainder of this article is structured as follows. First, we discuss ‘Background and Related Work’. Then, we elaborate on the ‘Method’ and describe the ‘Results’. We then propose a ‘Framework’ for prescriptive process monitoring methods, and conclude the paper in ‘Conclusion’.

Background and Related Work

Process mining takes event logs of a system that supports the execution of a business process to discover process models (Van Der Aalst, 2012). An event log contains information, such as timestamps, activity, unique case id, resources, and other contextual attributes about executed cases. Thus, process mining finds its applications in the discovery of data-driven process models which can then be analyzed and improved. Process mining is also used for conformance checking which helps to evaluate whether a real process corresponds to a process model and process enhancement through which a process model can be enriched with additional information (Van Der Aalst, 2012), such as performance data (Milani & Maggi, 2018). With the advancement of technology, process mining has developed beyond process discovery, conformance, and enhancement (Milani et al., 2022). A large number of predictive process monitoring techniques have been proposed that are able to predict time-, risk-, and cost-related outcomes, as well as sequences of outcomes and inter-case metrics (di Francescomarino et al., 2018). These predictions can also further be used for prescriptive process monitoring.

Methods for prescriptive process monitoring prescribe interventions that can change the outcomes of an ongoing process case. For instance, if a method detects that an undesired outcome is probable to unfold, an alarm is raised that can lead to an intervention (Teinemaa et al., 2018). This intervention could e.g., come in the form of an action performed by a process worker, such as calling a customer, that helps to mitigate or prevent the negative outcome from materializing (Fahrenkrog-Petersen et al., 2022). Thus, it is essential to define a policy for when a prescription is generated. The aforementioned example (Fahrenkrog-Petersen et al., 2022) e.g., considers the probability of a negative outcome and evaluates the cost model and mitigation effectiveness before triggering interventions. In other words, to decide whether an intervention should be prescribed, this particular technique first checks how probable it is that the ongoing case will end in a negative outcome. If the probability is high, it also calculates the trade-off between the cost and effect of intervening. Based on the outcome of the cost-effect analysis, this technique triggers an intervention.

Few prior studies focus on areas that are related to prescriptive process monitoring. For example, di Francescomarino et al. (2018) introduce a value-driven framework that allows companies to identify when to apply predictive process monitoring methods. The authors provide examples of predictive process monitoring techniques from the perspective of inputs that the techniques use, tools that implement them, and the domains where they had been tested. Another classification of predictive process monitoring methods has been proposed by Márquez-Chamorro, Resinas & Ruiz-Cortéz (2018), who focus on methods to train predictive models. In particular, the authors describe building and evaluating a predictive model. In Mertens et al. (2019), the authors evaluate predictive methods used to recommend follow-up activities in the healthcare domain. Thus, the focus there lies in describing the application of predictive process monitoring for processes prevalent in hospitals. While prescriptive process monitoring methods often incorporate predicted outputs, our work solely focuses on prescriptive methods and prescribed interventions.

In Lepenioti et al. (2020), the authors review methods for prescriptive analytics. The authors provide a taxonomy of existing prescriptive analytics methods and map out research challenges and opportunities. Pufahl et al. (2021) present a systematic literature review on automatic resource allocation. The authors provide an overview of approaches with regards to resource allocation goals and capabilities, use of models and data, and their maturity. Similarly, Arias et al. (2018) give an overview of resource allocation methods, but with a particular focus on human resources. However, Lepenioti et al. (2020) reviews prescriptive methods in general, and the latter two (Pufahl et al., 2021; Arias et al., 2018) focus on resource allocation. In this article, we build on such works by considering various types of potential interventions in process-aware methods.

Method

We conducted a systematic review of the existing body of work on prescriptive process monitoring methods. More specifically, what the objectives for using such methods are (RQ₁), what interventions such methods prescribe (RQ₂), what data such methods require (RQ₃), which modeling techniques they employed (RQ₄) and which policies (RQ₅) the methods use. The main objective of the SLR was, thus, to explore different aspects of prescriptive process monitoring. Therefore, we searched for methods and case studies that explore or study how ongoing process cases can be improved through the intervention of caseworkers. We use the method of SLR as it is particularly suitable for identifying relevant literature on a specific research topic (Kitchenham & Charters, 2007). We followed the guidelines proposed by Kitchenham & Charters (2007), which consist of three main steps: (1) planning the review, (2) conducting it, and (3) reporting the findings.

For the first step (planning), we identified research questions and developed the review protocol (Kitchenham & Charters, 2007). Then, we developed a search string for the review protocol, identified suitable electronic databases, and defined inclusion and exclusion criteria. Finally, we defined the data extraction strategy.

In the search string, we included “process mining” to scope the study to methods that rely on event logs. We derived the term “prescriptive” from the research questions. We also included the terms “recommender” (e.g., de Leoni, Dees & Reulink, 2020; Yang et al., 2017) and “decision support” (e.g., Mertens, 2020), as we found these to be sometimes used instead of “prescriptive”. Accordingly, we formulated the following search string:

While conducting the first search, we noted that the term “prescriptive process monitoring” was not consistently used. Therefore, only using this search string might have resulted in missing relevant studies. We addressed this issue by examining the papers we identified using the first search string to identify other terms. We noted that terms such as “next-step recommendation” (Huber, Fietta & Hof, 2015), “next best actions” (Weinzierl et al., 2020a), “proactive process adaptation” (Metzger, Kley & Palm, 2020) have been used synonymously for “prescriptive process monitoring”. We also noticed that the phrase “business process” often appeared in titles and keywords. Therefore, we formulated the following second search string:

We applied both search strings to the ACM Digital Library, Scopus (includes SpringerLink), the Web of Science, and IEEE Xplore to identify potentially relevant papers. The databases were selected based on the coverage of publications within the field of process mining. Finally, we conducted backward referencing (snowballing) (Okoli & Schabram, 2010) to identify additional relevant papers.

Next, we defined exclusion and inclusion criteria (see Table 1). For the set of exclusion criteria, we excluded papers for which the answer to any of the defined exclusion criteria was “no”: not digitally accessible (EC1), not in English (EC2), duplicates (EC3), and shorter than six pages (EC4). Exclusion criteria EC1 and EC2 ensured that the paper could be generally accessed and understood by other researchers. Papers that were unavailable via open access or via subscription services of the University, or via internet search were excluded. Papers in any other language than English were also removed. Criterion EC3 removed duplicates due to the paper being included in several digital libraries. We applied criterion EC4, as papers with less than six pages were not likely to contain sufficiently detailed information about the prescriptive process monitoring method for our analysis. We also defined three inclusion criteria: (IC1) the paper is relevant to the domain of prescriptive process monitoring, (IC2) the paper presents, reviews, discusses, or demonstrates a method or a case for prescriptive process monitoring, (IC3) the paper describes at least one way to identify candidate interventions for an ongoing process case. Thus, the answer for the inclusion criteria had to be “yes”. IC1 aimed to include papers that are inside of the prescriptive process monitoring domain. With IC2, we made sure to include studies that represent a theoretical discussion or practical application of a method. Inclusion criterion IC3 ensured that the papers contained sufficient information to address our research questions. Following a top-down approach, if a paper failed an inclusion criterion, it was excluded without the other criteria being considered.

Finally, we defined our data extraction strategy (see Table 2). We first captured the metadata of all papers (title, author, publication venue, year). Then we defined the data required to address the research questions. Thus, for RQ₁, we defined the data required to identify the objective of using the prescriptive process monitoring technique and performance metric(s) targeted in each paper. Next, we defined the data to elicit the interventions prescribed, the process perspective, and the users the prescribed interventions are presented to (RQ₂). Then, we defined the required data input for the techniques described in the different papers (RQ₃). Finally, we added modeling techniques (RQ₄) and policies used to trigger interventions (RQ₅) to the data extraction strategy.

We executed the search ¹ and identified a total of 1,367 papers (see Table 3). We filtered them using exclusion criteria EC1 and EC2. This resulted in the removal of 97 papers. Thus, 1,270 papers remained and were filtered based on EC3. Out of the remaining 1,010 papers, we removed short papers (EC4). This resulted in 900 papers remaining. These were filtered by title, thus removing papers that were clearly out of scope. The remaining 171 papers were filtered by abstract, resulting in 66 papers remaining. Finally, we applied the inclusion criteria by reading the whole paper and removed 44 papers. As a result, 22 papers remained. A total of 15 papers were added through backward referencing, resulting in the final list of 37 relevant papers.²

To derive the framework, we started by clustering the prescriptive process monitoring methods described in the identified papers according to what they were aiming to improve (RQ₁), e.g., “cycle time minimization”, “cost optimization”. For each group, we followed the research questions to classify the methods further, such as according to the interventions they trigger (RQ₂), the input data they require (RQ₃), the modeling techniques they use (RQ₄), and the policies they utilize to trigger interventions (RQ₅).

Results

In the following subsections, we present the results of our review. First, we provide a quantitative overview of the identified papers. Next, we describe the identified objectives of prescriptive process monitoring methods (RQ₁) followed by the interventions prescribed to achieve these objectives (RQ₂). Next, we outline the data (RQ₃) and modeling techniques (RQ₄) used. Finally, we describe the different policies used by the prescriptive methods to trigger interventions (RQ₅).

Prescriptive process monitoring objectives

In our review, we identified two main objectives that prescriptive process monitoring methods aim to achieve. The first objective is related to optimizing the process outcome whereas the second concerns optimizing the process efficiency. We adopt this classification based on the definition of quality and efficiency of service by Dumas et al. (2018). The objective of optimizing the process outcome relates to ensuring that the process outcome is positive. This objective is commonly expressed with binary metrics, such as avoiding a deadline violation (Gröger, Schwarz & Mitschang, 2014). The second objective of optimizing process efficiency relates to a particular quantitative aspect of process performance. Therefore, the objective of optimizing process performance can be expressed as, for instance, reducing cycle time (Wibisono et al., 2015).

As to the objective of optimizing the process outcome, five papers discuss undesired temporal outcomes, such as violation of a planned cycle time or deadline (Gröger, Schwarz & Mitschang, 2014; Sindhgatta, Ghose & Dam, 2016; Weinzierl et al., 2020a; Huber, Fietta & Hof, 2015; de Leoni, Dees & Reulink, 2020). For instance, Gröger, Schwarz & Mitschang (2014) describe an example from a manufacturing process, where the target is to reduce deadline violations. Thus, the possible outcomes can be described as binary: a deadline is violated (i) or a deadline is not violated (ii). Another set of studies focuses on avoiding or mitigating an undesired categorical outcome (Teinemaa et al., 2018; Fahrenkrog-Petersen et al., 2022; Metzger, Kley & Palm, 2020; Shoush & Dumas, 2021; Ghattas, Soffer & Peleg, 2014; Thomas, Kumar & Annappa, 2017; Mertens, 2020; Haisjackl & Weber, 2010). For example, Ghattas, Soffer & Peleg (2014) aim to avoid customers rejecting delivery in a bottle manufacturing process. In the domain of healthcare, examples of undesired outcomes are patients entering a critical stage (Thomas, Kumar & Annappa, 2017), or medical mistakes due to patient restrictions (Mertens, 2020).

The second main objective considers optimizing the process efficiency. Most papers consider optimizing temporal perspectives (15 out of 21), such as cycle or processing time. More specifically, in Wibisono et al. (2015); Kim, Obregon & Jung (2013); Obregon, Kim & Jung (2013); Thomas, Kumar & Annappa, (2017), reducing cycle time is defined as the main objective. In this context, reducing cycle time means that the cycle time should be gradually improved by ensuring that each coming case takes less time than the previous average. Thus, it differs from the previously described objective of avoiding deadline violation. For instance, Thomas, Kumar & Annappa, (2017) describe a method to minimize the cycle time of an environmental permit application process. Others focus on processing time, i.e., time spent by a resource resolving a task (Dumas et al., 2018). For instance, in Park & Song (2019) the aim is to reduce the processing time of manual tasks in a loan application process. Another set of methods aims at reducing the defect rate. For example, a method aims to reduce the likelihood and severity of a fault based on risk prediction (Conforti et al., 2015). Two papers (Goossens, Demewez & Hassani, 2018; Terragni & Hassani, 2019) describe methods that aim to improve revenues, e.g., by increasing customer lifetime value (Goossens, Demewez & Hassani, 2018). Finally, two methods (Khan et al., 2021; Schonenberg et al., 2008) can be utilized for different targets. As such, a method by Schonenberg et al. (2008) provides an example of optimizing the cycle time but the authors suggest that it could also be used to optimize costs, quality, or utilization.

Framework

Having reviewed the papers, we followed the research questions to identify the characteristics of the prescriptive process monitoring methods we identified. Starting from RQ₁, we clustered the methods according to their objective. We then followed the remaining research questions to detail the clustering. As a result, the proposed framework (Figs. 2 and 3) characterizes prescriptive process monitoring methods based on ten characteristics. The framework³ reads from left to right and starts with the objective of using a prescriptive process monitoring method. It then goes on to describe the interventions to reach that objective, the input the method requires, the techniques it uses, and the policy to trigger interventions.

Research Gaps and Implications

The presented framework provides an overview of existing prescriptive process monitoring methods by categorizing them according to their objectives and targets. The framework also presents different available methods and the different ways these methods enable reaching particular objectives. The overview, however, also unveils several gaps and associated implications for future research.

First, we observe that the majority of studies tested their methods with synthetic and/or real-life event logs. Therefore, the validation is done using a real-world or synthetic observational event log, but crucially it is not done in real-life settings. An attempt to test the effectiveness of interventions in real-life settings was made by Dees et al. (2019). Their study showed that predictions were rather accurate, but the interventions did not lead to desired outcomes. Thus, the proposed methods should be validated in real-life settings to ensure their usefulness in practice (Márquez-Chamorro, Resinas & Ruiz-Cortéz, 2018).

Second, our review showed that the majority of methods focus on identifying cases in which interventions should be applied and finding the point in time an intervention should be triggered during the execution of a case. In contrast, little attention has been paid to the problem of discovering which interventions could be applied to optimize a process with respect to a performance objective. Discrete methods leave it up to the users (stakeholders) to define possible intervention(s) a priori (e.g., de Leoni, Dees & Reulink, 2020; Teinemaa et al., 2018). Methods that use observational event logs from BPIC, rely on winner reports to identify possible interventions (e.g., Shoush & Dumas, 2021; Teinemaa et al., 2018). Continuous methods, in contrast, focus on recommending the next task(s) (e.g., Goossens, Demewez & Hassani, 2018; Yang et al., 2017) or resource allocation (e.g., Arias, Munoz-Gama & Sepúlveda, 2016; Abdulhameed et al., 2018; Kim, Obregon & Jung, 2013). Thus, one direction for further research could be to design methods that support the discovery of interventions from business process event logs, textual documentation, or other unstructured or structured process metadata.

Related to the above problem of discovering interventions, we observed that methods differ with regard to the nature of prescriptions. As our review shows, there is a body of methods that focus on providing prescriptions to guide a user during process execution (e.g., Terragni & Hassani, 2019; Weinzierl et al., 2020b). These methods are commonly based on the similarity of an ongoing case with previous executions of the same process. At the same time, there are methods that specifically optimize processes according to a performance measure (e.g., Khan et al., 2021; Metzger, Kley & Palm, 2020). Thus, methods differ between simply guiding the user and providing a specific suggestion for an action to be taken next. On this basis, future researchers can make a distinction for the type of method and, respectively, prescriptions they are developing. As such, the spectrum of options starts with guiding methods which require developing an algorithm to compare the ongoing case with the past executions which ended in desired outcomes to recommend the suitable option for continuing the running case. Another option is a category of methods that are optimizing. Such methods would typically require incorporating predictions based on past executions. One way is to consider the correlation of a predicted value with the likelihood of it having a positive effect on the running case. Another way is to focus on causality, where the impact of the proposed intervention on the outcome of the case should be estimated—for example, through building a causal estimation model (Shoush & Dumas, 2021; Bozorgi et al., 2021).

Another gap in existing research relates to the problem of designing and tuning policies for prescriptive process monitoring. Existing prediction-based methods (e.g., Sindhgatta, Ghose & Dam, 2016; Gröger, Schwarz & Mitschang, 2014) prescribe an intervention when the probability of a negative outcome exceeds a defined threshold. However, because predictive models are based on correlation (as opposed to causal relations), the prescriptions produced might not address the cause of a negative outcome or poor performance (e.g., the cause of delay). In this respect, we note that only a few existing methods take causality into account when designing policies (e.g., Shoush & Dumas, 2021; Bozorgi et al., 2021). Thus, developing policy design techniques that take causality into account can be another direction for further research.

As discussed in Dees et al. (2019) and de Leoni, Dees & Reulink (2020), the choice of whether or not to apply an intervention or the choice of which intervention to apply often depends on contextual factors. Some interventions may prove ineffective or counter-productive, for example, due to second-order effects (i.e., an intervention has a consequence that has a subsequent consequence). For example, an intervention wherein a customer is contacted pro-actively in order to prevent a complaint may actually increase the probability of a complaint (Dees et al., 2019). Similarly, assigning a resource to a case that is running late might lead to other cases being neglected, thus creating delays elsewhere and subsequently resulting in a higher ratio of delayed cases. Detecting such second-order effects requires human judgment and iterative policy validation (e.g., via A/B testing (Kohavi & Longbotham, 2017)). In this respect, it is striking that existing prescriptive process monitoring methods do not take into account the need to interact with human decision-makers. A crucial step in this direction is the ability to explain why a prescriptive monitoring system recommends a given intervention. There are two aspects to this question. First, explaining the prediction that underpins a given prescription (prediction explanation), and second, explaining the policy that is used to trigger a prescription (policy explanation). A possible direction to enhance the applicability of prescriptive monitoring methods in practice is to integrate explainability mechanisms. While several proposals have been made to enhance the explainability of predictive process monitoring methods (Hsieh, Moreira & Ouyang, 2021; Rizzi et al., 2022), the question of policy explanation in the area of prescriptive process monitoring is—to the best of our knowledge—unexplored. In other words, current methods do not incorporate mechanisms to explain why an intervention is recommended for a given case and in a given state.

Besides the aforementioned gaps, our findings also highlight that the majority of methods in the field of prescriptive process monitoring aim to improve processes along the temporal perspective (e.g., cycle time, processing time, deadline violations (Abdulhameed et al., 2018; Park & Song, 2019)). In comparison, other performance dimensions are only represented in a few cases (defect rate in Detro et al., 2020; Conforti et al., 2015; Yang et al., 2017, revenue in Goossens, Demewez & Hassani, 2018; Terragni & Hassani, 2019). Thus, another research direction could be to investigate other performance objectives that could be enhanced via prescriptive process monitoring.

Finally, our review also highlights a lack of common terminology in the field. This might be due to the novelty of the research field of prescriptive process monitoring. Authors use a large variety of labels. As such, the terms “proactive process adaptation” (Metzger, Kley & Palm, 2020), “on-the-fly resource allocation” (Wibisono et al., 2015), “next-step recommendation” (Huber, Fietta & Hof, 2015; Nezhad & Bartolini, 2011) are all used to describe the development and application of prescriptive process monitoring methods. This highlights the need for common terminology.

In summary, we note the following research gaps:

• Validating methods in real-life settings to ensure their usefulness in practice.

• Designing methods to support the discovery of interventions from event logs and assessing their potential effectiveness.

• Developing policy design techniques that take causality into account.

• Improving explainability of prescriptive process monitoring methods.

• Investigating other performance objectives rather than temporal that could be enhanced via prescriptive process monitoring.

• Developing a common terminology in the field.

Conclusion

Our article provides an overview of the research on prescriptive process monitoring and outlines a framework for categorizing methods in this field. The framework categorizes existing methods according to their objective, target metric, intervention type, technique, data input, and policy used to trigger interventions.

Our findings indicate that existing prescriptive process monitoring methods primarily have one of two objectives: optimizing process outcome or optimizing process efficiency. In order to achieve the respective objective, a range of interventions can be prescribed. Our review indicates that the interventions commonly concern control flow and resource perspectives. We also note that interventions can be divided into two categories: intervention frequency (when interventions are prescribed) and intervention purpose (how interventions are prescribed). According to the intervention frequency characteristic, the corresponding methods can be continuous (prescribing an intervention for multiple or all activities of an ongoing case) or discrete (an intervention is triggered only when a need is detected). As to intervention purpose, the interventions can be optimizing or guiding. On the one hand, optimizing methods suggest interventions to improve an ongoing case with respect to a certain KPI or performance measure. Such interventions can be correlation- or causality-based. On the other hand, guiding methods provide recommendations solely based on an analysis of historical traces.

Our review also shows that existing prescriptive process monitoring methods use control flow, resource, temporal, and domain-specific data as required input. This data is processed by modeling techniques which, as we found, differ based on whether a method is optimizing or guiding. Similarly, the policies to prescribe an intervention differ based on the method. As such, guiding methods use similarity-based policies, while for optimizing methods the policies can differ (e.g., set of rules, maximum metric improvement, probability of a negative outcome above a threshold, etc.).

We also identified research gaps and associated research avenues to highlight where the field of prescriptive process monitoring is headed (i.e., answering the question “Quo vadis?”). In particular, our work highlighted: (i) a lack of in vivo validation of the proposed methods; (ii) a lack of methods for discovering suitable interventions and assessing their potential effectiveness; (iii) little emphasis on explainability and feedback loops between a prescriptive monitoring system and its end-users; and (iv) a narrow focus on temporal metrics and comparatively little work on applying prescriptive monitoring to other performance dimensions. While our results might not be surprising for process mining experts, they provide an overview of the current state of the art in the field of prescriptive process monitoring and give insights for future research directions.