A Survey of Process-Oriented Data Science and Analytics for supporting Business Process Management

Our paper aims to conduct a systematic literature review by analysing research studies related to process analytics. Our goal is to understand the recent developments in the field of process analytics by examining the following research questions:

1. RQ1 What is the body of recent and relevant academic publications within the field of business process analytics?

2. RQ2 What family of methods exists for diagnostic analytics that lets us mine relevant process data in a retrospective (post-mortem) fashion?

3. RQ3 What aspects of business processes can be predicted?

4. RQ4 What are some of the state-of-art methods for predictive monitoring tasks?

5. RQ5 What role can prescriptive process analytic methods play in automating or supporting decision making in the context of business process execution?

6. RQ6 How do we characterise these findings in a taxonomy?

7. RQ7 What should be the future research focus of BPM and process analytics?

RQ1 is the core question that identities existing methods to support the practice of business process management. It allows us to identify a set of classification criteria. Given the richness of the process analytics literature, this overarching question is then decomposed into subsequent research questions to have a well-delimited and manageable scope. RQ2 aims to identify methods that attempt to understand the process behaviour using offline data in order to diagnose possible performance problems and identify areas of potential improvements. Given the vast number of publications in the broader field of process mining, we focus on contemporary topics that have not been examined in great detail so far. RQ3 and RQ4 investigate aspects of business processes that can be predicted by means of machine learning techniques (also known as business process monitoring techniques). RQ5 explores prescriptive methods with the potential of providing decision support both at strategic and operational level.

We categorize our findings based on input data required, type of algorithm employed, validation method, and tool support. (Note to self: The strategy for taxonomy isnt finalized yet).

Keeping the research goals in mind, and following the guidance given in (Kitchenham, 2004) we define the search string. We drive a set of relevant keywords from our subject matter expertise:

• “business process” — generic term for retrieving most process analytics papers.

• “mining” - Here a relevant study must take as input a event dataset and proposes a technique for analysing, extracting actionable knowledge.

• “diagnostic” — diagnostic methods perform post-mortem analysis of process data in an attempt to understand what happened in the past and to analyse the root causes of performance issues.

• “monitoring“ — monitoring methods are concerned with run-time predictions of process outcomes.

• “prediction” — a relevant study that estimates or predicts various aspects/properties of a business process.

• “prescriptive” - studies that are concerned with action recommendations

• “decision-support” - Methods that provide decision support for achieving process goals.

• “recommendations” - Methods that provide recommendations for decision support.

• •

  “process mining AND performance improvement”

• •

  “process mining AND resource management”

• •

  “business process prediction”

• •

  “predictive business process monitoring”

• •

  “prescriptive process analytics AND decision Support”

• •

  “prescriptive process analytics”

• •

  “process recommendations AND decision-support”

• •

  “decision support AND process analytics”

• •

  “decision support AND knowledge-intensive processes”

We then picked the Google Scholar database (Gusenbauer, 2019) as our primary search engine for retrieving relevant studies based on our final search strings. We double-checked the retrieved studies with six other literature search sources, including SpringerLink, Scopus, IEEE Xplore, ScienceDirect and ACM Digital Library. These major database electronic literature databases cover the scientific publications within the field of computer science. After search was conducted(in August 2020), it returned more than 5000 papers.

IN1: We picked the study only if: (i) it is concerned with analysis and mining process data. (ii) proposes techniques for predictive monitoring of processes. (iii) investigates methods related to prescriptive analytics and decision support in the context of structured, unstructured and knowledge-intensive business processes.

IN2: Study is published in 2011 and later were all included, even if they had fewer than 5 citations. However, for papers published before 2011 we used the snowballing technique.

IN3: The study clearly defined research context, goals and proposes a novel method that has been properly evaluated with sound experiments.

IN4: The study is peer-reviewed and published in a reputable venue.

For final inclusion, a given study must meet the above inclusion criteria. We applied the inclusion criteria IN2 by configuring the search engine’s filter settings. For applying the IN1 and IN3 criterion, was assessed by reading title, abstract and skimming the relevant sections of the paper. After filtering the search results by applying inclusion and exclusion criteria, we identified primary and subsumed studies. Primary studies constitute an original contribution to the field and subsumed are ones that do not substantially improvement with respect to the original contribution. The application of the exclusion criteria resulted in 734 relevant studies out of 1319 works selected in the previous step.

We categorize the selected works using different dimensions specifying the typology of the existing methods and their characteristics.

In particular, each study can be decomposed and organized along the following dimensions:
- Input data
- Outcome
- Process perspective (control flow, resources, data)
- Family of algorithms(the main algorithm used in the study)
- Evaluation data (real-life or artificial logs) and application domain (e.g., insurance, banking, healthcare)
- Implementation (standalone or plug-in, and tool accessibility)

Lastly, we also observed that currently, many of the same mining problems are being studied under different names, and proposed methods are being evaluated in an ad-hoc manner with no common experimental setups or evaluation measures.

Prior to process design, organizations often engage in an exercise of strategic planning by using balanced scorecards where the mission and strategy of an organization (Goodspeed, 2004) are translated into a set of performance measures (also known as KPIs). Such performance measures, can help organization keep track of progress towards the specific defined organizational goals. Moreover, for organizations implementing business process management, organizational objectives are used to formulate process performance measures, characterized by metrics like per-instance cost, cycle time efficiency (CTE), resource utilization, quality of service and so on (Zur Muehlen and Shapiro, 2015). Several performance measurement systems exist, and appropriate ones can be picked depending on the organizational objectives or strategic success factors (Vom Brocke and Rosemann, 2010). Devil’s quadrangle can for example, measures dimensions such as process cost, quality (e.g. visiting frequencies, error rates) and cycle time (also known as throughput time)(Jansen-Vullers et al., 2007). Jansen-Vullers et. al (Jansen-Vullers et al., 2007) have proposed a framework for quantifying the impact of best practices and discuss different dimensions for performance measurement. Several domain-specific performance reference models also exist. e.g Supply Chain Operations reference model, IT infrastructure Library (ITIL) etc.

Performance analysis techniques allow us to extract, analyze and enhance existing process models by identifying performance improvement areas and mining KPI values using the available process logs. Early identification of large deviations between the planned and actual KPI values allows organisations to take corrective actions to achieve the desired goals. Performance analysis techniques include Bottleneck Analysis (where we identify activity, resource and waiting bottlenecks), workload and demand analysis (where we analyze resource usage to identify under-utilized or over-utilized resources), rework analysis (where we identify errors or defects) and over-processing analysis. Organizations can leverage these techniques to identify a range of issues related time, costs and quality based on event log data. These insights can later be used in the process redesign phase for improving or enhancing the apriori process model. We note that process performance analysis has several titles in the literature. e.g Business Process perspective, Performance perspective, Process Performance Management etc (Hornix, 2007) (Zur Muehlen and Shapiro, 2015).

Performance analysis techniques can be divided into Qualitative or Quantitative analysis (Dumas et al., 2013a). In Qualitative Process Analysis, we can perform Value-Added Analysis(which aims at identifying waste) and Root Cause Analysis(e.g cause-effect analysis and why–why analysis). Here we measure process performance using metrics like per-instance cost, cycle time efficiency (CTE), resource utilization, quality of service, and so on (Zur Muehlen and Shapiro, 2015). Quantitative process analysis is based on historical process execution data or simulation models and involves using techniques like flow analysis, queuing theory and process simulation (Dumas et al., 2013a). Process simulation is a popular technique for quantitative analysis of process models where synthetic data based on hypothetical instances is used to estimate cycle times, average waiting times and average resource utilization. Similarly, flow analysis provides a set of techniques that utilize activity performance data to estimate the process’s overall performance but ignores the resources utilization aspect from the analysis (Dumas et al., 2013a).

Business process monitoring deals with event analysis and support of real-time executions of a given process. It aims to understand the present process performance by presenting a real-time picture of key performance indicators(often using dashboards) such as mean execution time, resource utilization or error rate etc.(Dumas et al., 2013b). Process monitoring tools help determine how well business processes under consideration are aligned with the goals of an organizations. They enable continuous, real-time monitoring of processes based on performance-related KPI values of a given process instance to track various business objectives. Process monitoring tools like performance dashboards can monitor the actual KPI values(e.g the distribution of cycles times, visualize the activities and execution patterns, and total duration) of a business process(Tardío and Peral, 2015). Several software vendors offer various performance analysis features where we can find the causes of these deviations(when they happen) and project performance analysis results onto process models to show bottlenecks, service performance levels, throughput times, and frequencies (Hornix, 2007).

To inform decision making under uncertainty, Businesses often draw on previous experience and recorded data to understand the potential downstream effects of certain specific decisions and actions. Historically, to determine the effect of a given intervention on a particular process would require either guesswork or launching numerous A/B tests, which can be costly and time consuming (Kohavi and Longbotham, 2017). Theory of causality provides a better alternative for potential downstream effects of decisions that are being considered (Pearl et al., 2009). Causal process mining seeks to use the process execution logs to discover and quantify cause-effect relations. Causal Process Mining can answer the fundamental question: What changes, if implemented, will cause an improvement to the process?. Existing process discovery techniques allow us to discover correlation but not causation. In causal analysis we try to develop an understanding that goes beyond the control-flow perspective. We are interested in understanding the outcomes(desired or undesired effects) associated with particular action(s) (interventions) taken during the course of process execution. There has been much recent interest in studying Causal analysis and Causal reasoning in the data science and machine learning research communities.

Leveraging process data would allows us to infer future process state based on what happened in the past and use that evidence to establish causality. However, pinning down causal effects rigorously is challenging. We briefly cover some of the recent contributions to process analytics in this section. Koorn et. al. (Koorn et al., 2020) have proposed a method that uses statistical tests to discover action-response-effect patterns. Their method can identify causal relations between responses (interventions) and effects for certain pre-defined sub-populations. This can support the decision making processes by giving insights into how certain actions lead to desired outcomes (e.g improved performance). In (Mahnaz Sadat Qafari, 2020) Qafari et. al. showed that root cause analysis using structural equation modeling is useful for testing if a predetermined causal relation (identified earlier by a process analyst) holds. Bozorgi et. al. (Dasht Bozorgi et al., 2020) proposed an approach that extracts recommendation rules from event logs. Causal effect can then be assessed and rules with highest incremental effect (uplift) can be used as recommendations in the form of interventions that can influence process outcomes positively. Their framework also computes a cost-benefit model of a particular intervention, identifying particular cases to which it is applicable (case-level recommendations). Agarwal et. al (Narendra et al., 2019) take a similar approach for structural causal model discovery and performing counterfactual reasoning. Discovering cause-effect relationships in event data is useful for root cause analayis of process performance issues but remains a challenging open problem. Hompes et. al. (Hompes et al., 2017) propose a method which uses Granger causality to generate a graph of causal factors explaining factors (or their combinations) that may affect process performance.

Causal Analysis techniques has also been used to understand process deviation, explain predictions and make recommendations. We will discuss these techniques in next sections.

Business process variant (also known as deviance analysis or drift detection) refers to process instances that maybe deviate from the desired course of execution, resulting in an unexpected or unplanned outcome. Variations may occur due to contextual/environmental factors, human factors or because of explicit decisions made by process participants. In a given process log, process variants are a subset of instances that violate the behavior prescribed by the model and can distinguished based on a certain characteristics they are correlated with. e.g. representing violation compliance rules or missing set performance targets. In deviance mining, we are interested in identifying various variants of the processes that may exist and diagnosing the root-causes of process variations by analyzing or comparing two or more event logs. i.e Given a set of event logs of two or more process variants, how can we identify and explain the differences among these variants? (Taymouri et al., 2021a). Understanding the factors leading to variation can help managers make better decisions and improve overall process performance.

Several methods have been proposed in the last decade to analyze the execution logs to identify and explain differences between two or more process variants. One class of techniques uses frequent pattern mining which typically takes as input two event logs (corresponding to two variants of a process) and produces as output a list of differences (with respect to established performance objectives). Machine Learning techniques on the other hand can classify instances as “normal” or “deviant” when trained with labelled examples. In (Nguyen et al., 2016) Nguyen et al. have surveyed sequence classification and mining techniques. They divide the techniques into those based on frequent pattern mining and rest based on discriminative pattern mining. They also provide benchmark comparisons of representative techniques using various real-life event logs. Similarly, Taymouri et al. (Taymouri et al., 2021a) have tried to present a unified view of variant analysis techniques by surveying and classifying existing methods based on data, type of algorithms and analysis performed. Several methods rely on identifying frequent patterns, while others use generative approaches to discover and compare models of process variants. They have categorized process variant analysis outcomes as Rule-based, model based or descriptive (representing discrepancies and behavior of the different process variants. Modeling and management of process variants is another important challenge that contemporary BPM tools do not adequately support (Reichert et al., 2015). Existing tools require that variants be specified in separate process models or ’expressed in terms of conditional branches within the same process model’.

Operational Decision-making associated with frequently executed processes and cases can significantly impact process outcomes. Traditionally, a business process model specifies activities within which the decision-making occurs. Decision logic can define the specific logic used to make individual decisions, such as business rules or executable analytic models. Decision modelling can represent complex, multi-criteria business rules and thus provides another perspective, creating a bridge between business process models and decision logic models (Figl et al., 2018).

Decision Model and Notation (DMN) is a modeling notation for decisions, published by OMG and allows for decoupling between decisions and control flow logic (Model, 2016). It provides an understandable symbolic representation of operational decisions and supports supports decision management, and business rules specification (Kluza et al., 2019).

Decision mining aims at mining the Decision model along with along with associated decision logic. The Decision model will define the decisions to be made in tasks (defined by the process model), their interrelationships, and their requirements for decision logic. Process execution data, depicting underlying rules (governing the choices) can be mined for frequently made decisions within an organization. However, decision mining is challenging due to several factors (e.g incompleteness of available data). Decisions are often dependent on personal expertise and contextual factors, information that may not be present in event logs. Decision mining techniques can identify and derive decision rules from relationships between process context, path decisions, and process outcomes. Several techniques proposed in the literature has tried to tackle the problem of decision mining partially. In mannhardt et al. (Mannhardt et al., 2016) propose a decision-tree based learning method to discover overlapping decision rules from event data. Similar to decision mining, Batoulis et. al. (Batoulis et al., 2015) have proposed a method for extracting decision logic from process models. The paper explains a semi-automatic approach where execution logs are not required and decision points can be replaced by generating a dedicated decision model, allowing decision logic to modeled separately from process logic. Rozinat et. al. (Rozinat and van der Aalst, 2006) propose a method for decision point analysis that determines how data attribute values(or data dependencies) affect the routing of a case. Their technique can identify decision points in Petri net models and has been implemented as a ProM Plug-in. Leoni et. al (De Leoni and van der Aalst, 2013) extended the technique and proposed a general-purpose method for discovering Branching Conditions (where atoms are equalities or inequalities consisting of multiple variables and arithmetic operators). Similarly, Bazhenova et. al. (Bazhenova et al., 2016) use decision tree classification to derive of DMN based decision models. Their techniques are able to extract not only control flow decisions but data decisions and dependencies. Overall, a comprehensive set of methods and frameworks is required that assist orgnaizations with dynamic management of decisions via analyzing, modelling and improvement efforts. Such methods should be able to leverage available process data and mine Decision model along with along with decision logic and dependencies between decisions and data elements .

Human resources often take the role of knowledge workers and play a major role in modern organizations where processes are increasingly complex have knowledge-intensive nature (Di Ciccio et al., 2015). In such knowledge-intensive scenarios, Resources (especially human resources) play a critical role in deciding the overall process outcomes. Event logs contain rich information about the task, resource and process outcome and can be leveraged to optimally realize the process goals (Rajan, 2018). Resource analysis (also known as organizational or resource mining/perspective) aims at analyzing event logs for extracting insights about organizational resources, evaluate resource performance and understand past resource allocation decisions in order to find areas of improvement (Pika et al., 2017). Over the past decade, various resource analysis techniques have been proposed to tackle problems like organizational structure discovery, classification of users in roles, discovery of organizational models, social network analysis (SNA), resource allocation, analysis of information flows between organizational entities and role mining (Zhao and Zhao, 2014) (Burattin et al., 2013) (Song and Van der Aalst, 2008)(Rajan, 2018).

Optimal allocation of resources to process tasks can significantly impact the overall process performance. Comparing and tracking the productivity of human resources is a challenging problem due to biases and lack of objectivity. Sindhgatta et al. (Sindhgatta et al., 2015) propose a method that uses process execution logs to identify resource allocations decisions which result in good outcomes(measured in terms of quality of service). In a similar work, Sindhgatta et al. (Sindhgatta et al., 2014) investigate the variation in resource efficiencies with varying case attributes and show that process outcomes are dependent on various contextual factors like complexity of work, task priority and capabilities(expertise) of the resources involved.

Accurately measuring resource behaviour allows us to effectively dispatching and suggest staffing policies that meet the contractual service levels (quality) of the service system and the business process. Huang et. al. (Huang et al., 2012) consider the problem of measuring resource behaviour from different perspectives such as preference, availability, competence and cooperation. Similarly, several methods for extracting useful knowledge about resource performance from event logs have been proposed in the literature. Pika et al. (Pika et al., 2017) propose a technique based on data envelopment analysis for analysing resource productivity using event logs. This technique is often used to measure the efficiency of companies. Linh ly et. al. (Ly et al., 2005) study the problem of mining staff assignment rules from event-based data using an organisational model. Similarly, Senderovich et. al. (Senderovich et al., 2014) explore mining of resource scheduling protocols from recorded event data. Related to the problem of resource assignment and allocation Cabanillas et. al. (Cabanillas et al., 2013) have studied the challenge of resource ranking and prioritization.

The success of process outcomes also depends on resources interaction. Social collaboration patterns between human resources can significantly impact process outcomes as optimal cooperative resource behaviour leads to enhanced service quality and process performance. Improving process performance by analyzing relationships and mining collaboration patterns between organizational entities is one of the key goals of resource analysis. It is based on the premise that process output not only depends on capability but compatibility as well. Social network analysis (Van der Aalst and Song, 2004) is one approach for understanding organizational structure and analyze relationships between originators involved in processes. Such an analysis is useful for many reasons such as improving handover relations, improving resource compatibility etc. In (Kumar et al., 2013) Kumar et al. propose a modeling technique that captures the compatibility between resources at the time of task assignment. Schonig et. al. (Schönig et al., 2015) proposed an approach to extract declarative process models. Their technique allows modelling of organizational relations using rule templates that can be represented in textual Declarative Process Intermediate Language (DPIL).

RBAC (Role based access control) is used to manage resources in workflow area and is a useful model for managing resources(Kuhlmann et al., 2003). Event logs can be used to to mine role based access control (RBAC) models, identifying the privileges or authorization of resources (Baumgrass, 2011) (Burattin et al., 2013). Roles refer to the assignment of organizational agents to job functions within an organization. Discovering an optimal set of roles remains a useful problem to investigate. Vaidya et. al. (Vaidya et al., 2007) define the perform as determining a role-based access control (RBAC) configuration and analyze its theoretical bounds. Similarly, frank et. al. (Frank et al., 2013) propose a probabilistic solution to learn the RBAC configuration and show how it generalizes well to new system users for a diverse range of data.

Modern organizations routinely deploy process analytics, including process discovery techniques on their process data, both to gain insight into the reality of their operational processes and also to identify process improvement opportunities (Augusto et al., 2018). Process analytics techniques such as process discovery play an important role in mining event data and providing organizations with insights about the behaviour of their deployed processes. However, in many practical settings, process log data is often geographically dispersed, may contain information that may be deemed sensitive and may be subject to compliance obligations that prevent this data from being transmitted to sites distinct to the site where the data was generated. Traditional process mining techniques operate by assuming that all relevant available process data is available in a single repository. However, anonymising, giving control access and safely transferring sensitive data across organization/site boundaries while preserving priacy guarantees is non-trivial. However, in many practical settings, process log data is geographically dispersed and can contain information that may be deemed sensitive. A classical example of a privacy-preserving process mining problem of the first type is from the field of medical research involving impediments to data migration. Consider the case that a number of different hospitals wish to jointly mine their process logs for the purpose of medical research, but are faced with regulatory and legislative compliance hurdles that prevent clinical process histories being shared across health jurisdictions (hospitals, health districts, national boundaries etc.). Hospitals are therefore restricted from ever pooling their data or revealing it to each other leading to small dataset available for knowledge extraction. This negatively impacts the confidence with which clinicians might deploy the results thus obtained. Our inability to migrate clinical process data also implies that we miss out on the opportunities for extracting higher-impact insights that might have been possible if data from multiple health jurisdictions could have been analysed in juxtaposition (Lenz and Reichert, 2007).

Traditional process mining techniques operate by assuming that all relevant available process data has been curated into a central site for analysis. However, anonymising, giving control access and safely transferring sensitive data across organizations is non-trivial. Moreover, organizations face legal constraints, risk of data breaches (or hacks) along with data integration challenges, preventing them from building a centralised data warehouse (Dunkl et al., 2011). This leads to a scenario where event-log data is present in organizational silos and distributed among several custodians, none of whom are allowed to share/transfer their sensitive data directly with each other (Lang et al., 2008). Mining process data in such cross-silo settings can prove to be invaluable for providing relevant operational support to organizations if privacy guarantees can be offered (Jensen et al., 2012).

Differential Privacy provides us with a formal privacy notion for datasets that are released publicly or might come in contact with potentially malicious adversaries (McSherry and Talwar, 2007). It is considered as the de facto standard for ensuring privacy in a variety of domains. The definition proposed by Dwork et al. (Dwork et al., 2014) offers a mathematically rigorous gold standard for ensuring privacy protection when analyzing datasets like process logs(or results of a randomized algorithm) that might contain sensitive or private information. We modify the definition slightly for event logs:

Definition 1: Differential Privacy (adapted from (Dwork et al., 2014)) A randomized mechanism $\mathcal{M}:\mathcal{D}\rightarrow\mathcal{R}$ with a domain $\mathcal{D}(e.g.,$, possible event logs) and range $\mathcal{R}(e.g.$, all possible trained models $)$ satisfies $(\epsilon,\delta)-$ differential privacy if for any two adjacent process logs $l,l^{\prime}\in\mathcal{D}$ and for any subset of outputs $S\subseteq\mathcal{R}$ it holds that $\operatorname{Pr}[\mathcal{M}(d)\in S]\leq e^{\epsilon}\operatorname{Pr}\left[\mathcal{M}\left(d^{\prime}\right)\in S\right]+\delta$

Two process log $l$ and $l^{\prime}$ are defined to be adjacent if $l^{\prime}$ can be constructed by adding or removing a single instance(entry) from the log $l$. By bounding the potential worst-case information loss, the above definition provides us with a strong formal privacy guarantee. Formally, under the $(\varepsilon,\delta)$-differential privacy definition, we measure Differential Privacy properties of our method by epsilon and delta values. Epsilon($\epsilon$) is the privacy loss parameter in differential privacy and is inversely proportional to the amount of noise added. i.e Lower values of $\varepsilon$ imply stronger privacy guarantees. A Differentially private mechanism typically involves using a randomized mechanim that perturbs the input dataset, intermediate calculations, or the outputs of a function, using a calculated quantity of noise (usually at the cost of utility) (Dwork et al., 2006). Such a mechanism is considered private if it hides the isolated contribution of any single individual in the databases. i.e removing a single entry will not result in much difference in the output distribution (Acs and Castelluccia, 2012) (Dwork et al., 2014).

In Privacy-Preserving Distributed Process Discovery, our goal is to discover a global process model by privately mining multiple distributed process log independently and share only the resulting insights from each analysis. i.e mining a differentially private process model, without ever pooling the data to a central site, in a way that reveals nothing but the final discovery process model to the participating organizations.

Most typical methods presented in the literature rely on some form of data transformation in order preserve user privacy. These techniques are a trade-off between information loss and privacy(Bamiah et al., 2012). Techniques based on Secure Sum Protocol (Clifton et al., 2002) permits a network of nodes to transmit a numeric sum (from node to node) to which each node adds a node-specific number without having any node being able to compute what the individual contributions of the participating nodes were. The final node obtains the sum of the numbers contributed by all of the participating nodes, again without being able to compute what the individual nodespecific numbers were. This protocol can be used to create distributed versions of most process mining algorithms. Elkoumy et al. (Elkoumy et al., 2020) propose an architecture based on Sharemind, which uses multiparty compute to mine a directly follows graph. However they don’t provide differential privacy guarantees.

In process mining, while a lot of emphasis has been on analyzing and extracting process insights from the observed behaviour logged in event logs, the knowledge dimension associated with business processes has received very little attention (Di Ciccio et al., 2015). i.e Current process mining techniques are self-contained and have minimal capacity to leverage and reasoning using prior knowledge. In practice, this means process mining algorithms can’t perform inference that goes beyond the implicit knowledge which is recorded in the event logs. Traditional mining techniques focus on mining behaviour that inherently cannot represent all of the cascading hierarchical structure representing complex real-world processes (Van der Aalst, 2009). These algorithms can’t reason about abstract relationships between various objects involved in the process. For example, easily-drawn inferences that people can readily answer without direct training like smoke is seen so there must be fire happening cannot be inferred by the current process mining methods. This leads to an incomplete understanding of process behaviour where process analysts are left trying to abstract, simplify, and even leave out key relationships needed for complete understanding of process behaviour.

Process knowledge has many faces and it will differ from other forms of organizational knowledge as it will be highly contextual, sometimes tacit and relevant to a particular domain. Organizations have realised that knowledge and processes are interlinked and should explicitly be made a key component of business processes (Barclay and Murray, 1997). Knowledge is more complex than simple data or information with an additional characteristic of being subjective, as its often tacit in nature (e.g obtained with years of experience). Process knowledge has many faces and it will differ from other forms of organizational knowledge as it will be highly contextual, sometimes tacit and relevant to a particular domain. Sometimes this knowledge is documented but even then it exists in silos of unstructured documents, often halting the progress of processes to consult any knowledge bases that might exist.

For organizations, having access to the right knowledge at the right time is crucial to effective decision making as it allows companies to solve problems quickly, diffuse best practices amongst employees, cross fertilize ideas and enable them to stay competitive (Van Beveren, 2002). Managing Process knowledge requires a deliberate and systematic approach which should be reflected at all levels of organization. Organizations employ various knowledge managements tools to capture and create the knowledge that is either explicit or tacit by sometimes interviewing experts, telling war stores, and apprenticeship style training programs. Despite several attempts, modern organizations find capturing managing the available knowledge(in terms of capturing, codifying, and sharing) difficult and rely on knowledge workers for know-how. This leads to organizations routinely forgetting the lessons learned during the past when knowledge workers leave or switch projects. Overtime organizations have realised the importance of deliberate and systematic approach for creating a culture where company’s knowledge base in maintained (Von Rosing et al., 2014) and many expert systems, and Enterprise knowledge management systems have been proposed to capture knowledge effectively, categorize it and then make that knowledge available across an organization. Such tools can codify knowledge in the form of wikis, cognitive maps, decision trees, knowledge graphs etc. These movements were largely unsuccessful(Easterby-Smith and Lyles, 2011).

Knowledge management has largely been ignored in recent years. Solutions are needed that tackle knowledge management challenges and help organization extract actionable knowledge from all available process related data(e.g extracting meaningful knowledge from unstructured documents) and make it widely available inside the organisation. Similarly, Common-sense reasoning has been highlighted as one of the major challenges for the process analytics research community (Calvanese et al., 2021). It follows a broader trend in AI research where the need for solving complex tasks by incorporating knowledge and common-sense reasoning has been repeatedly highlighted (Davis and Marcus, 2015). Furthermore incorporating domain knowledge by means of constraints can support process analysts in their efforts to fully understand executed process behaviour recorded in real-world event logs and improve the outcome of process mining activities.

The topics covered in this section, along with classic topics like process discovery continue to provide opportunities for research contribution. We note that even with the availability of various state of the art process discovery methods, understanding business processes and resource behaviour from process logs alone remains a challenging problem(Benatallah et al., 2016). Furthermore systematic adoption of process mining in challenging domains like healthcare, also remains a challenge (Munoz-Gama et al., 2022). Process mining techniques can be of great value to answer process-related questions and is often a starting point for process analysts, however, more support is needed to address the challenges faced by process analysts. Carmona (Carmona, 2020) has highlighted some of the open problems and research challenges associated with process discovery and conformance. Their findings show that the existing mining methods struggle to deal with challenges like spaghetti models, concept drift and identifying events that occurred at different levels of abstraction. We also observe that traditional process mining and analysis techniques have not given significant attention to the analysis of resource behaviour and its affects on process outcomes. Resource analysis deserves more attention from the research community as knowledge about resource behaviour can be used for effective planning, strategizing and gaining insights which lead to better overall process performance. In (vom Brocke et al., 2021) Brocke et al. present an enterprise framework for analyzing the effects of process mining that emerge at various levels of an organization. Lastly, by understanding contextual factors and gaining detailed insights about how processes are being executed within a particular context also remains an interesting challenge.

Data Challenges: Process mining techniques are primarily reliant on process logs, which don’t always explicitly capture all the behaviour of past executed processes(Augusto et al., 2018). Such logs are susceptible to domain gaps, data bias (due to incompleteness) and quality issues (due to noise and erroneous data recordings). Many real world processes are unstructured in nature and for these processes most state-of-the-art process discovery algorithms produce, hard-to-interpret, spaghetti-like models which poorly fit the event log. This negatively influences the usefulness of the discovered process model. Often times discovered models are hard to interpret from a process analyst perspective while also prone to under-fitting or over-fitting the given event logs, offering only minuscule support for improving process outcomes. Furthermore, a particular challenge in process mining is the management of business-process variants and contemporary business process management tools do not provide adequate support for modeling and management of process variants (Taymouri et al., 2021b). For complex domains like healthcare, where improving clinical outcomes can directly impact the quality of life for patients, this implies that process analysts miss out on the opportunities for a complete understanding of the underlying process behaviour and subsequently extracting higher-impact insights.

Event data can come from various heterogeneous data sources and in several data formats. These logs are rarely in the desired format and form. It is common for process analysts to apply a series for pre-processing steps for consolidation, verification(checking for errors and inconsistencies) and transformation. To identify an event trace representative of a process instance, is challenging and is often done manually using techniques like extraction, correlation, and abstraction of the event data by the process analyst (Diba et al., 2020). There is also the challenge of ’Big data’, where data to be analyzed is of huge volume, velocity and variety, making it unfeasible to be analysed with traditional analysis tools. Other challenges include the diversity of formats and nonstandard data models (Sakr et al., 2018). To maximize the utility of available data, it is sometimes useful to build a data lake that provides a single consolidated view of organizations’ datasets. Data lake makes the preparation and analysis(by querying) of data more accessible. Such data can also include relevant context in which business processes were executed (this could be the context for an entire process or an individual task) and enable organizations to gain insights that help improve existing processes.

As observed in various other fields of AI and machine learning, an important driver of measuring progress is precisely defining Process Analytic tasks(using mathematically rigorous definitions) and building useful benchmark datasets that can be used for performance comparison (Blagec et al., 2020). The progress of the field depends on the cycle of identifying real-world problems, researchers proposing novel techniques and industry adaptation of such techniques. It is therefore important to make quantifiable progress by benchmarking process analytics tasks on standard datasets with well-defined performance metrics. We observe that many process mining published papers often lack strong reproducible experiments, resulting in poor comparability and relative merits of the proposed approach. So far, only in predictive process analytics, we have seen the adoption of rigorous evaluation criteria. We refer to work by (Augusto et al., 2018) as an example of such an effort where Augusto et al. review various state-of the art process discovery methods and benchmark the performance of various automated discovery algorithms. We hope future studies will follow the same approach where they critique and compare proposed methods against existing state-of-art process mining techniques via standard benchmarks.

Next Activity Prediction refers to the problem of predicting the next trace suffix likely to occur during the execution. In Next Activity Prediction, we attempt to derive a process-agnostic machinery for learning to generate recommendations and predict the future behaviour of a given partially executed process instance while assuming minimal domain knowledge. Models are trained using historic data and after training, the input is using the observed prefixes of running process cases (event stream). The problem has been tackled using various machine learning techniques where it is often treated as a multi-class classification problem. The next event(of a process instance) can be represented as one class that the ML classifier can predict.

Prediction techniques based on deep learning have a popular choice and have shown promising results(Evermann et al., 2017) (Márquez-Chamorro et al., 2017). Such techniques are often motivated by the applications of deep learning to Natural Language Processing tasks (e.g langauge modeling). Tama and Comuzi (Tama and Comuzzi, 2019) and Weinzierl et. al. (Weinzierl et al., 2020b) provide a comparison of architectures and encoding techniques for next activity prediction using real world logs. We can use several evaluation metrics(e.g Accuracy, F1-score etc.) to compare the methods. Brunk et al. (Brunk et al., 2020) consider the problem of context-sensitive process predictions(in business process monitoring) by employing evidence sensitivity analysis to determine if context is cause or effect of the next event during execution. This allow us to understand the impact that a context variable can have on a running instance and offers an explanation for understanding why a certain prediction was made. While LSTMs based techniques can theoretically deal with long event sequences, the long-term dependencies between distant events in a process get diffused into the memory vector. We therefore seek modeling methods that are more expressive and allow storing and retrieval of intermediate process states in a long-term memory. To tackle this, recently, Khan et al.(Khan et al., 2018) explore the application specific type of neural network known as the memory–augmented neural network (MANN) for the task of next event. In a typical setting, a MANN is a recurrent neural network (e.g., LSTM [8]) augmented with an external memory matrix. Differential Neural Computer architecture can be adapted the to account for a variety of tasks in predictive process analytics: (i) separating the encoding phase and decoding phase, resulting dual controllers, one for each phase; (ii) implementing a write-protected policy for the memory during the decoding phase.

Predicting the evolution of running cases is an important problem and plays a key role in risk management, resource allocation and process improvement. Similar to Next Activity, Process Path Prediction methods can predict possible paths of a running instance up to its completion(Verenich, 2016). Tax et. al. (Tax et al., 2017) and Evermann et al. (Evermann et al., 2016) explore the use of Long Short-Term Memory (LSTM) to solve various process predictive problems. Tax. et al.(Tax et al., 2017) show the use of LSTM neural networks for predicting the sequence of the future activities. Similarly, Camargo et al. (Camargo et al., 2019) employ deep learning techniques to predict sequences of next events, their timestamp, and their associated resource pools. Considering context(as cause or effect) when making these predictions is also an interesting problem studied by Brunk et. al. (Brunk et al., 2020). Tama and Comuzzi provide a comprehensive empirical comparison and bechmarking for such techniques (Tama and Comuzzi, 2019).

Time-related Predictions problems such as Predicting the remaining cycle time, ompletion time and case duration, predicting delayed process executions and deadline violations of running instances has been studied extensively. Indicators available in event logs can be exploited to make predictions about time-related risks(Van der Aalst et al., 2011). Verenich et. al. (Verenich et al., 2019b) have done an extensive survey of various methods used for predicting the remaining cycle time and present a cross-benchmark comparison. Van Dongen et. al. (van Dongen et al., 2008) apply non-parametric regression on event data to predict remaining cycle time. Similarly (Evermann et al., 2017) employ RNN to predict the duration of activities. Similarly, Tax. et al.(Tax et al., 2017) propose a prediction method based on Long Short-Term Memory (LSTM) neural networks.

The problem of case outcome prediction aims at identifying process instances that will end up in an undesirable state (measured as likelihood and severity of fault occurrence or violation of compliance rules) (Verenich, 2016). Instances can be labelled as normal or deviant and then process related risks can be classified using any of the traditional modern machine learning techniques. The case outcomes can be assessed primarily by first checking if the process has met its ’hard goals’ and then soft goals (determined by KPIs such as time quality, cost etc.) Similarly, in Compliance monitoring techniques are aimed at preventing Compliance violations by monitoring ongoing executions of a process and checking if they comply with respect to certain business constraints(Verenich, 2016).

Explainability Explainablility or Interpretability remains a key challenge in predictive business process monitoring. Interpretability is defined as ”the ability to explain or to present in understandable terms to a human” and often stems from incompleteness in problem formulation leading to unquantified bias (Doshi-Velez and Kim, 2017). Business process stakeholders and decision makers can only fully trust prescriptive support systems that offer an explanation for the decisions or recommendations made (Adadi and Berrada, 2018). Decision Support systems which leverage machine learning methods must therefore be able to explain the reasoning behind certain decisions, recommendations, predictions made or actions taken. Predictive process monitoring methods should explain why the predictive model was mistaken when the predictions are inaccurate. By making systems and models interpretable we want to ensure that decisions and data can be explained to end users in a transparent and easy to understand manner (Adadi and Berrada, 2018). Increasingly, Explainability and Interpretability is not just a desirable property but increasingly becoming a serious matter of public debate. In some countries compliance laws will require ’a right to explanation’ which means end-users can ask for an explanation of a certain decision that affects their lives (Goodman and Flaxman, 2017). This also means that use of black box methods (such as Deep Neural Networks) in predictive and prescriptive approaches will be infeasible as we can’t explain the decisions or predictions made. ’Explainablility’ and ’establishing trust’ in these models remains one of the key challenges, not only for sensitive domains like healthcare but now for everyday consumer-facing products as well. Therefore, developing systems that provide trustworthy explanations and the necessary chain of reasoning that led to particular decisions and outcomes remains a significant challenge for future research work in predictive and prescriptive monitoring. Several methods have been proposed to make systems more comprehensible. Rizzi et. al. (Rizzi et al., 2020) for example propose a method that uses post-hoc explainers and encoding for identifying the most common features that explains incorrect predictions. By reducing the impact of identified features, explanations can be used to improve model accuracy. Brunk et. al. (Brunk et al., 2020) take a different approach to the problem of making transparent context-sensitive predictions by proposing a next event prediction technique based on dynamic Bayesian network. Galanti et. al. (Galanti et al., 2020) propose a based framework based on game theory of Shapley Values for explaining the predictions. Their framework is based based on LSTM models, and can explain any generic KPI. Lastly, Verenich et. al. (Verenich et al., 2019a) employ flow analysis technique for predicting quantitative performance indicators of running process instances. Their technique can be used to estimate values of this performance indicator (e.g cycle time) by aggregating performance indicators of the activities composing the process. We refer the readers to the work by Doshi-Velez and Kim (Doshi-Velez and Kim, 2017) and Adadi and Berrada (Adadi and Berrada, 2018) for a more detailed discussion around the topics of ’Interpretability’ and ’Explainable AI’.

Modern organizations operate in dynamic environments and face the challenge of dealing with uncertainties(risks) associated with various process decisions due to environmental factors that are often not fully under their control. Managing process-related risks by making risk-informed decisions remains a key challenge for organizations. Risk reduction usually involves decreasing the likelihood and severity of faults during process execution and ensuring that desired performance goals are met.

Management of risks during process execution remains a challenging problem(Suriadi et al., 2014). The ability to detect potential metric deviations early on is valuable in giving process owners decision information for a timely intervention. One primary goal of process decision support during Business process execution is to guide risk-informed decisions at run-time by modeling, detecting and mitigating risks as early as possible. In practice, this means using predictive and prescriptive techniques to identify process instances that are likely to get delayed or to terminate abnormally. This involves using those predictions to intervene early, recommend actions or support resource allocations decisions at run-time that enable organizations to avoid or decrease the likelihood of metric overrun.

To tackle these challenges, Risk-based decision support (RDS)(Sometimes known as Risk-aware Business Process Management Systems) have been proposed to help with early detection of factors by predicting risks in terms of metrics deviations during process execution. They can make recommendations that can help minimize the predicted process risk and avoid negative outcomes. The performance criteria of business processes are typically described by Key Performance Indicators (KPIs), and their target values can be defined based on business goals. They can assist in risk management by monitoring KPIs, PPMs, and QoS metrics. Overall, reduce risks (in terms of metrics deviations during process execution) to provide decision support for a given process, e.g., recommending the next process activity, which minimizes process risks. When those deviations exceed a tolerance threshold, interventions can be taken to decrease the likelihood of unfavourable outcomes.

Risk reduction techniques allow us to identify process instances that are at risk of not meeting certain performance criteria and recommend preventive actions to process participants. Several techniques have been proposed in the literature for modeling and detection of risk. Suriadi et. al. (Suriadi et al., 2014). Provide a comprehensive review of techniques for managing process-related risks. Literature on prescriptive business process monitoring consists of techniques (Gröger et al., 2014) (Conforti et al., 2013a) (Schonenberg et al., 2008) that can be used to recommend preventive actions in order to support risk-informed decision making.

Future Business Process Management systems must, therefore, assist process users/stakeholders in risk-informed decision making by generating predictions and recommendations that help reduce such risks (Di Francescomarino et al., 2017). Risk Management and BPM were historically separate fields, and their integration is understudied, leaving room for future research contributions (Suriadi et al., 2014). Conforti et al. (Conforti et al., 2015) discuss the Risk-aware BPM lifecycle where each phase can be complemented with elements of risk management like Risk Identification, Risk-aware Execution and Risk monitoring etc. Lastly, problems like real-time risk detection, resource scheduling, Automated risk mitigation and Real-Time Risk Monitoring are also investigated under the umbrellas of risk-informed decision making(Conforti et al., 2011) (Conforti et al., 2012) (Conforti et al., 2013b). Overall, managing process-related risks by making risk-informed decisions remains a key challenge for organizations.

In the previous section, we discussed how predictive monitoring techniques can provide operational support based on models learned using historic logs in order to predict what is likely going to happen next and use those prediction capabilities to influences the outcomes of running case instances. e.g. monitor processes and issue recommendations to workers and managers based on probability that a given case will violate the set performance targets. The output of Predictive business process monitoring techniques, is just predictions. Predictions can be used as early warnings for taking risk informed decisions but do not explicitly support answering of question like What action should we take next to achieve a particular goal? and Why should we do it?(Lepenioti et al., 2020). Compared to descriptive and predictive business analytics, prescriptive process analytics remains less mature (Eili et al., 2021). Marquez et. al. (Márquez-Chamorro et al., 2017) point out that ‘little attention has been given to providing recommendations’. Instead of providing specific action recommendations, literature on business process monitoring focuses on forecasting future process events(and outcomes) while leaving the action implementation part to the subjective judgment of process users and business decision makers(Dees et al., 2019).

Recommender Systems have found applications in information filtering system such as in video or music services as playlist generators or content recommendations for social media companies etc(Beheshti et al., 2020).

Process-aware Recommender Systems have been proposed to assist knowledge workers, in operational decision-making by recommending actions for executing a particular process/task, manage resource allocation policies and so on (Beheshti et al., 2020) (Schonenberg et al., 2008). Eili et al. (Eili et al., 2021) provide a systematic review of Recommender Systems in Process Mining and classify recommendation approaches as ‘pattern optimization’, ‘risk minimization’, or ‘metric-based’. For structured processes, Process-aware Recommender Systems have been proposed to assist knowledge workers, in a context-aware adaptable fashion by recommending actions for executing a particular process/task, manage resource allocation policies and so on (Beheshti et al., 2020) (Schonenberg et al., 2008). Such systems leverage technologies like machine learning to build recommender systems that monitor process instances, predict future process states and recommend appropriate actions(Dees et al., 2019).

We should note that most process-aware recommender techniques focus on supporting risk-informed decision making.Their major focus is on preventive measures early warning recommendations to for example avoid predicted metric deviation. Another major goal of Process-aware Recommender systems should be to support decisions that that maximize the likelihood of achieving business goals. e.g Weinzier et al. (Weinzierl et al., 2020a) consider problem of recommending next best actions that lead to optimal outcomes. Their technique relies on explicitly adding control-flow knowledge to their proposed technique via formal process model and uses process simulations to verify and filter the predictions of the trained predictive model. Groger et al. (Gröger et al., 2014) introduce the concept of recommendation-based business process optimization to support adaptive process execution. Their framework recommends actions for the next process step to take for a given process instance. For organizations data-driven process optimization enables decision support for real-time process optimization. e.g. shortening the reaction time of decision-makers to events that may affect changes in process performance.

Business Processes assist organizations in organizing activities that deliver business value, usually in the form of a product or a service. Over the last decade, automation has caused the landscape of work to change significantly and Knowledge workers are now regarded as the most valuable organizational assets. Knowledge work is characterized by unstructured processes which can be hard to specify at design-time Supporting knowledge workers involved in the execution of unstructured Knowledge-Intensive Processes by providing context-specific recommendations remains an interesting challenge.

Knowledge Intensive Processes (KIPs) are processes that require precise expert(tacit) knowledge, involvement of knowledge workers, and consisting of activities that do not have the same level of repeatability as structured processes(Di Ciccio et al., 2015). Instead of assuming a rigid process structure, knowledge-intensive processes (KiPs) are goal-oriented, often unstructured(with pre-defined fragments) and characterized by activities that cannot be anticipated(or modeled in advance). KIPs represent a shift from the traditional process management view (where process models are structured with repeatable tasks), to a model where task execution depends on knowledge workers as primary process participants (Di Ciccio et al., 2015).

Knowledge workers are highly trained and have specialized expertise in performing complex tasks autonomously and are considered a key asset for modern businesses. Supporting knowledge work when rigid definitions of process models are not available or cannot be designed apriori (with structured or unstructured data)remains a key challenge(Di Ciccio et al., 2012). They typically rely on their experience based intuition and domain expertise, for decision-making. Their work is less characterized by explicit procedures and more by creative thinking that usually cannot be planned a priori(Di Ciccio et al., 2015). An example of knowledge work is Clinical decision-making for patient treatment in the hospital emergency room, which is highly case-specific and requires a knowledge-driven approach. i.e Treatment Decisions are made based on highly specific medical domain knowledge, the context in the form of patient’s medical history, years of specialized experience and evidence that emerges from patient test results and real-time sensors.

Knowledge workers still lack the adequate decision support tools to assist them in executing knowledge-intensive processes(Di Ciccio et al., 2015). As we enter the knowledge economy, future process management systems will have to drive processes in modern enterprises that are highly knowledge-driven, are semi-structured or unstructured while leveraging a diverse range of process-related datasets. e.g recommend appropriate actions to knowledge workers while operating in dynamic environments. In (Di Ciccio et al., 2015) Ciccio et al. have provided a set of requirements for process-oriented systems to support knowledge-intensive processes. One of the inherent challenges that future process management systems must address is that of flexibility. Flexibility means instance-specific adaptations based on context and environment. Flexible execution of business process instances involves multiple critical decisions at each step. e.g. what task to perform next and what resources to allocate to a task and so on. Previously, to address the problem of supporting flexible knowledge-intensive process, many paradigms have been proposed. e.g Adaptive Process Management(ACM), Flexible Process Aware Information Systems, case handling systems and declarative processes. Adaptive Case Management gained the most popularity amongst the various paradigms.

Adaptive Case Management(ACM) is aimed at supporting knowledge workers involved in the execution of dynamic, unstructured knowledge intensive processes(KIPs) where course of action for the fulfillment of process goals is highly uncertain(Hauder et al., [n.d.]) (Motahari-Nezhad and Swenson, [n.d.]). ACM offers a way to manage the entire life-cycle of a “case” by following the ’planning-by-doing’ principle, where work is done by considering the context and is continually adapted based on the changing characteristics of the environment(Motahari-Nezhad and Swenson, [n.d.]). In the case management paradigm, the focus is on the case and its hard to pre-define the sequence of activities. For example, Case is the ‘Product’ being manufactured or a ‘patient’ being treated where primary driver of case progress is the case data and information that emerges as the case evolves. There however can exist template or patterns that represent the structured aspects of the process. Here process could be seen as a recipe for handling cases of specific type (van der Aalst et al., 2003). A Case template is created by the knowledge worker, allowing high degree of flexibility in executing a particular case. Templates can then be used to instantiate case instances and represents a middle ground between a completely specified structured process and an unstructured process. Case execution allows us to gather feedback and adapt the templates to be reused in a particular context(Marin et al., 2016). Adaptive Case Management (ACM) has been gaining significant interest for handling unpredictable situations in processes and still lacks ’common operational semantics’ and a ’proper theory’ (Hauder et al., [n.d.])(Hewelt and Weske, 2016).

In the context of Case Management, the problem of recommending ’next best steps’ in a case management system based on the knowledge of past similar cases(which is the focus of this work) has been addressed by Schonenberg et al. (Schonenberg et al., 2008) and Motahari-Nezhad et al. (Motahari-Nezhad and Bartolini, 2011). Schonenberg et al. attempt it by first finding similar cases based on abstraction, then using support and Trace Weights to consider the relative importance of a log trace. Similarly, Motahari-Nezhad et al. (Motahari-Nezhad and Bartolini, 2011) have looked at the problem of decision support for guiding case resolution based on how similar cases were resolved in the past. Such support can complement knowledge workers decisions, often made using personal experience and expert knowledge(Di Ciccio et al., 2015). We argue that the notion of recommendations under-pins decision support for not only structured processes but across the whole spectrum of process management. i.e we see recommendations as a general-purpose mechanism for providing operational decision-support for not only structured but semi-structured and unstructured processes as well. In knowledge-intensive processes, for example, recommendations can provide intelligent assistance to process users by offering concrete support in various process related decisions like resource allocation decisions or action recommendations etc. Such form of assistance allows knowledge workers to take preemptive actions that can avoid negative outcomes (Schonenberg et al., 2008)(Gröger et al., 2014). Data-centric AI approaches hold the promise of supporting knowledge intensive processes and case management practices, whereby enabling flexible process execution. Khan et al. (Khan et al., 2021) propose a data-driven reinforcement learning based recommender system for supporting knowledge workers that considers the past execution data in addition to characteristics of the objects involved (e.g product or user). The proposed system recommends the next best action (or sequence of actions) while taking into account asset characteristics and process context.

Resources are entities that are responsible for performing activities of a business process and must satisfy various (sometimes contradictory) business goals. Management of resources involves selecting the right resources, evaluating the efficiency of resources and optimally allocating tasks to relevant resources (Rajan, 2018). Decision support for resource management provides process users, intelligent assistance on the optimal allocation of resources based on their capabilities, past performance, current workload and process characteristics.

Past execution data containing resource allocation decisions and can be leveraged to provide real-time decision support regarding the allocation of resources at an appropriate time while considering the specific context. Taking this data-driven approach improve productivity and efficiency of business processes and helps avoid performance deviations. Selecting the most appropriate resource and assigning it to the right task is often a challenging decision as it is dependent on task complexity, task priority and expertise of the worker. Traditional resource allocation decisions are made based on profile matching(perceived capabilities), which often involves human judgment. Such approaches are not always optimal because of various challenges such as resource unavailability, overloading and uncertainty(of process execution and resource behaviors) (Sindhgatta et al., 2015). Traditional thinking also held that the performance of a process is determined by its design; thus, well-designed careflows would lead to better patient outcomes. More generally, though, process context also plays an important role. Process context can be defined as knowledge exogenous to a process, and not consumed as input to a process that nevertheless serves as a determinant of process performance and proposed a context-aware recommender system for identifying context and using it to support resource allocation and task allocation decisions.

A number of methods related to learning, reasoning, and planning resource allocation have been proposed in the literature. e.g. In (Rajan, 2018) Rajan et al. have explored addressed the question of context-aware process management. Such recommender systems can derive data-driven business process provisioning that supports effective dispatching and staffing policies and assist managers in meeting the desired quality of service (or performance) levels. Russel et al. (Russell et al., 2005) discuss workflow resource patterns and have identified three main allocation types, namely, capability-based allocation(where we match capabilities of available resources with and requirements of an activity), history-based allocation (using past execution data to make ) and finally Role-based allocation(which considers the organizational position and relation of the resource). Liu et al. (Liu et al., 2014) model the task allocation problem as a Markov Decision Processes (MDPs) and show their Q-Learning based method overcomes many of the shortcomings of traditional methods(e.g. load imbalance) to compute social relation between two resources. We can also consider various abstraction levels when allocating resources. e.g. Arias et al. (Arias et al., 2016) have proposed a recommendation system that dynamically allocates resources at a sub-process level based on multi-factor criteria (to assess resources). Their proposed tool considers metric scores in the various dimensions of the resource process cube(knowledge base) to present a ranked list of suitable resources. Similarly, machine learning approaches can be used to mine resource allocation rules. e.g. Huang et. al. (Huang et al., 2011) treat resource allocation as a sequential decision making optimization problem and propose a reinforcement learning based, resource allocation solution. Their Q-learning based framework allows adjustment of real-time allocation decisions by learning appropriate allocation policies based on available feedback.

Goal-orchestration for flexible process execution: Goal models holds the promise of delivering significant value Strategic Modelling, by providing a hierarchic representation of statements of stakeholder intent, with goals higher in the hierarchy (parent goals) related to goals lower in the hierarchy (sub-goals) via AND- or OR-refinement links. Goal models encode important knowledge about feasible, available alternatives for realizing stakeholder intent represented at varying levels of abstraction. A number of prominent frameworks leverage goal models, including KAOS, i* and Tropos (Wang and Han, 2004). Strategic Modelling for organizations selecting amongst alternative goal refinements (OR-refinements). Given a goal model that delimits that space of goals and sub-goals that an organization can seek to satisfy, this is a critical (and indeed, only) decision problem to be solved. An AND-refinement of a goal is a statement of know-how that tells the organization how to achieve a parent goal (although without sequencing information, and thus falling short of being a full procedure or process model). OR-refinements offer alternative specifications of know-how for a given parent goal.

Replacing tasks or activities with goals in-process models allows us to enact processes in, flexible, context-sensitive ways. In (Santipuri et al., 2017) Santipuri et al. introduce the concept of goal orchestration for modeling processes (or process behavior) to enable flexible process execution. Goal orchestrations offer abstract, strategy-level views on processes, which can aid human understanding and ease process redesign. Their proposed technique can mine goal orchestrations from enterprise event logs and compute alternative task-level realizations of a goal if the initial attempt at realizing the goal fails to achieve the desired results. Goal-oriented process mining is a promising sub-field. In (Ghasemi and Amyot, 2020) Ghasemi et al. present a survey various techniques for mining goals from event logs and argue that combining goals and process mining can potentially augment the precision, rationality and interpretability of mined models.

Achieving resilience in adversarial settings: The need to future-proof businesses is widely acknowledged as one of the hardest challenges facing business decision-makers. Businesses need to anticipate environmental changes and the likely behavior of competitors. Much of what happens in the business environment (the effects of moves by these actors) is adversarial in nature and adversarial moves prevent or impede the achievement of business goals. Decision-making in adversarial settings involves reasoning about chains of moves and counter-moves by the adversarial entities involved. Strategic resilience requires that businesses make decisions that are most resilient to adversarial moves by players in the business environment. e.g Red teaming allows businesses to informally reason about the adversary’s strategic planning process(Hoffman, 2017). Similarly, in (Gou et al., 2017) Gou et al. present a decision support framework for robust process enactment. They leverage adversarial game search (e.g Monte Carlo game tree search) to compute alternative flows in order to anticipate and account for possible ways in which the execution environment might impede a process from achieving its desired effects or outcomes. This notion can be extended for strategic decision-making as well, where decision-making is viewed through an abstraction of a two-player game. Here goal models (representing structured models of strategy) can be combined with game-tree search using augmented game trees to assist organizations in selecting amongst alternative goal refinements (OR-refinements). The final computational machinery, overall, provides business management with a resilient strategic decision-making framework that can reason the consequences of various decisions.