Source-linked AI summary
Predictive Process Monitoring Methods: Which One Suits Me Best?
Chiara Di Francescomarino, Chiara Ghidini, Fabrizio Maria Maggi, Fredrik Milani
TL;DR
Companies face a growing and difficult-to-navigate body of predictive process monitoring methods when deciding what their data can support and which techniques to use. The paper systematically reviews and categorizes this research, then develops a value-driven framework for comparing methods by predictions, data, algorithms, validation, and tool support. The framework is intended to guide organizations toward methods fitting their needs, while its presented version is abridged and its validation on real company users remains future work.
Problem
Companies need help navigating the growing range of predictive process monitoring techniques to determine what can be predicted from their data and which methods to use.
Method
The paper conducts a systematic literature review and extracts prediction, data, algorithm, validation, and tool-support information to build a value-driven framework.
Results
The resulting framework presents predictive monitoring options alongside required input data, tools, validation domains, algorithm families, and supporting research.
Takeaways & Limitations
Organizations can use the framework to identify predictive process monitoring methods that fit their needs and evaluate their applicability and potential benefits.
Takeaways & Limitations
The article presents an abridged framework, while empirical evaluation with real company users is planned for future work.
Abstract
from arXiv · showhide
Predictive process monitoring has recently gained traction in academia and is maturing also in companies. However, with the growing body of research, it might be daunting for companies to navigate in this domain in order to find, provided certain data, what can be predicted and what methods to use. The main objective of this paper is developing a value-driven framework for classifying existing work on predictive process monitoring. This objective is achieved by systematically identifying, categorizing, and analyzing existing approaches for predictive process monitoring. The review is then used to develop a value-driven framework that can support organizations to navigate in the predictive process monitoring field and help them to find value and exploit the opportunities enabled by these analysis techniques.
1 Introduction
Predictive process monitoring is emerging as a data-driven business enabler, but the growing range of techniques makes it difficult for companies to determine what to predict and which methods fit their data. The paper addresses this challenge through a systematic review and a value-driven classification framework.
- Predictive process monitoring uses growing operational data availability and maturing approaches to extend companies’ data-driven insights.
- Companies may struggle to navigate the many predictive techniques and identify suitable predictions and methods for their available data.
- The paper develops a value-driven framework for classifying existing predictive process monitoring research.
- A systematic literature review identifies, categorizes, and analyzes existing methods before informing the framework.
2 Background
Process mining relies on event logs containing ordered case events and associated attributes. Predictive process monitoring uses historical traces to forecast ongoing executions, extending reactive compliance monitoring toward preventive action.
- Process mining starts from event logs whose ordered events describe individual process cases and may include resource, timestamp, event, and case attributes.
- Predictive process monitoring belongs to process mining because it retrieves predictive information from historical execution traces.
- Traditional compliance monitoring identifies violations after they occur, whereas predictive monitoring provides forward-looking information about ongoing executions.
- Advance knowledge of violations, deviations, or delays can support preventive measures such as reallocating resources and avoiding money loss.
3 Systematic Review Protocol
The review follows a structured protocol to identify relevant predictive process monitoring research and extract comparable information about predictions, data, algorithms, validation, and tools. It narrows an initial database search to a final set of 51 papers.
- The review protocol specifies research questions, search procedures, selection criteria, and data extraction in two researcher-designed phases.
- The main research question asks what relevant academic publications exist, with subquestions covering prediction type, input data, algorithms, validation, and tool support.
- The search combines predictive and business-process keywords and covers six computer-science databases.
- From the final 51 papers, the authors extracted metadata, prediction types, required input data, algorithm families, validation settings, and tool support.
4 Predictive Process Monitoring Dimensions
Predictive process monitoring is organized by prediction outputs and algorithm families, with methods spanning numeric, categorical, and future-activity sequence predictions. The reviewed approaches differ in model structure, inputs, validation contexts, and operational targets such as time, cost, risk, and constraint outcomes.
- The framework classifies methods by prediction type and algorithm type, while the review also records input data, validation, and tool support.
- Prediction types: Prediction outputs comprise numeric measures, categorical or boolean outcomes, and sequences of future activities with their payloads.
- Algorithm families: Methods generally separate training from runtime prediction, then use either explicit models or machine-learning and statistical models built from encoded event-log information.
- Numeric predictions: Numeric predictions include time and cost, with time methods estimating completion or remaining time through transition systems, sequence trees, regression, clustering, or queueing approaches.
- Numeric predictions: Some numeric approaches incorporate concurrent-case status, and the resulting encodings improved results in two real-life case studies.
- Categorical predictions: Categorical predictions cover risks and categorical outcomes, using approaches that may rely on explicit process models, machine learning, constraint satisfaction, or statistical risk indicators.
- Categorical predictions: Predicate-fulfillment prediction can use an ongoing case’s event sequence and latest activity payload, although the cited framework incurs high runtime overhead.
5 Value-driven Framework for Selecting Predictive Monitoring Algorithm
The value-driven framework organizes predictive process monitoring methods by prediction type and practical selection criteria, helping companies identify approaches suited to their data and context.
- Prediction types: The framework classifies predictive process monitoring algorithms into six high-level prediction types, including time prediction.Time prediction covers execution-time aspects such as remaining time and delay.
- Input data: Companies begin selection by matching the desired prediction type with the required input data, including event logs and, for outcome predictions, a labelling function.Event logs may contain timestamps and additional data; outcome-based predictions usually require the predicate or category to be predicted.
- Tool support: The framework records tool support and distinguishes stand-alone applications from research-framework plugins such as ProM plug-ins.Tool support makes techniques easier to use, evaluate, and understand in terms of applicability and potential benefits.
- Validation: Validation information includes log type and industry domain because real-life and same-domain logs provide more relevant evidence for assessing algorithm suitability.Synthetic-log validation is described as weaker because it may not mirror the complexity and variability of real-life logs.
- Algorithm family: The framework identifies each method’s foundational algorithm family, such as regression, neural networks, or queuing theory, rather than listing the specific algorithm.This algorithm-family information is intended to support suitability assessment.
- Framework use: Reading the framework across its columns links prediction type, predicted aspect, required inputs, available tools, validation domain, algorithm family, and supporting research.The complete framework contains additional data, while the article presents an abridged version because of space limitations.
6 Conclusion
Predictive process monitoring research and techniques have grown rapidly, creating selection challenges for companies. The paper addresses this through a systematic literature review and a framework intended to guide technique selection, while proposing future empirical evaluation with real company users.
- Conclusion: Predictive process monitoring approaches have grown rapidly, but their expanding complexity makes it difficult for companies to navigate the field.The paper frames this difficulty as a consequence of the spread of available techniques.
- Conclusion: A systematic literature review is used to provide companies with a framework for selecting techniques that fit their needs.The framework is presented as guidance for technique selection.
- Conclusion: The proposed framework is planned for future empirical evaluation with real company users to assess its usefulness in real contexts.This evaluation is identified as future work.