Source-linked AI summary
A systematic literature review on state-of-the-art deep learning methods for process prediction
Dominic A. Neu, Johannes Lahann, Peter Fettke
TL;DR
Process-prediction research lacks comparable evidence because deep learning approaches vary in preprocessing, architectures, targets, datasets, metrics, and baselines. This paper conducts a systematic review of these approaches, classifies their design choices, and concludes that qualitative synthesis is possible but reliable performance ranking is not.
Problem
Deep learning process-prediction studies use divergent designs and evaluation settings, making their results difficult to compare quantitatively.
Method
The paper conducts a systematic literature review and classifies deep learning approaches by neural-network type, prediction type, input features, and encoding methods.
Results
The review compares 32 different approaches and identifies three broad architecture families and three prediction-target types.
Takeaways & Limitations
The review provides a qualitative comparison of design decisions and identifies conflicting statements and research gaps for further research.
Takeaways & Limitations
Sparse overlap in testing data and evaluation methods prevents a reliable performance ranking across the reviewed approaches.
Abstract
from arXiv · showhide
Process mining enables the reconstruction and evaluation of business processes based on digital traces in IT systems. An increasingly important technique in this context is process prediction. Given a sequence of events of an ongoing trace, process prediction allows forecasting upcoming events or performance measurements. In recent years, multiple process prediction approaches have been proposed, applying different data processing schemes and prediction algorithms. This study focuses on deep learning algorithms since they seem to outperform their machine learning alternatives consistently. Whilst having a common learning algorithm, they use different data preprocessing techniques, implement a variety of network topologies and focus on various goals such as outcome prediction, time prediction or control-flow prediction. Additionally, the set of log-data, evaluation metrics and baselines used by the authors diverge, making the results hard to compare. This paper attempts to synthesise the advantages and disadvantages of the procedural decisions in these approaches by conducting a systematic literature review.
1 Introduction
Process prediction extends process mining from retrospective analysis toward runtime decision support by forecasting the future of ongoing process instances. This review synthesizes diverse deep learning approaches whose preprocessing, architectures, goals, datasets, metrics, and baselines make results difficult to compare.
- Process models support descriptive and retrospective analysis but not operational decision support during execution.
- Predictive process monitoring forecasts the future of individual ongoing instances, enabling organizations to adapt execution and prevent undesired outcomes.
- Deep learning approaches gained popularity because multilayer networks learn complex features without handcrafted feature engineering.
- Existing deep learning process-prediction implementations differ in preprocessing, architectures, prediction goals, datasets, evaluation metrics, and baselines, limiting comparability.
- The paper conducts a systematic literature review, classifies identified approaches, and highlights conflicting statements and research gaps.
2 Related Work
Prior reviews address process prediction broadly or focus on particular prediction types, but differ in scope, methodology, and quantitative evaluation. This study positions itself as a structured, comprehensive review centered on deep learning process prediction and detailed design decisions.
- Earlier reviews analyzed process-prediction methods using process awareness, prediction targets, input data, tools, and application criteria.
- Outcome-oriented and remaining-time reviews focused on specific prediction types, while their coverage of deep learning was limited.
- The review examines neural-network architecture design at finer granularity than work primarily comparing process-aware and non-process-aware methods.
- A prior deep learning survey compared only three approaches and did not conduct a systematic literature review.
- Unlike related work, this study uses defined inclusion and exclusion criteria to provide a structured and comprehensive review of deep learning-based process prediction.
3 Background
Process prediction uses event-log traces to forecast future case behavior during execution, while deep learning supplies alternative neural architectures and encodings for sequence data. Applying these methods requires converting event attributes into numerical representations suited to the selected network.
- Deep learning: Deep learning uses multilayer neural networks, while feed-forward, convolutional, and recurrent architectures implement different computation patterns.
- Process prediction: Event logs record process instances as event sequences with trace identifiers, activities, timestamps, resources, and additional payload attributes.
- Process prediction: Process prediction is an online task that forecasts potential outcomes of incomplete traces during execution rather than analyzing only finished instances.
- Process prediction: Prediction separates realized prefix events from unknown suffix events, because only the realized events are available at prediction time.
- Deep learning based process prediction: Early deep learning process-prediction work targeted next-event and outcome prediction from sequences of observed activities.
- Deep learning based process prediction: Activity sequences and additional attributes must be numerically encoded according to whether they are categorical, numerical, or temporal before entering the neural network.
4 Methodology
The review uses a Kitchenham-based systematic literature review to identify and synthesize deep learning approaches for process prediction. It defines research questions, searches major computer-science databases, applies eligibility criteria, and supplements database results with forward searching.
- Review design: The review follows Kitchenham et al.’s methodology, beginning with research-question definition, search-string construction, eligibility filtering, and subsequent analysis.Its goal is to synthesize advantages and disadvantages of procedural decisions in deep learning for process prediction.
- Research questions: Six research questions address neural-network type, preprocessing, encoding, prediction target, dominant approaches, and combinability.The first four examine the state of the art, while the final two orient future research.
- Search strategy: The search covered Web of Science, ScienceDirect, IEEE Xplore, ACM Digital Library, and SpringerLink using domain, prediction, and deep-learning or neural-network terms.Terms were searched in titles, keywords, and abstracts, with adapted strings used to reduce false positives.
- Eligibility criteria: Eligibility required an implemented and evaluated business-process-prediction approach using deep learning, while excluding non-English, inaccessible, and out-of-scope studies.The inclusion criteria also covered neural networks and combinations of traditional machine learning with deep-learning architectures.
- Study selection: 21 papers remained after title-and-abstract screening and full analysis, while a forward search added 11 studies for a total investigation corpus of 32.The initial screening rejected 38 papers, and in-depth analysis rejected 19 additional studies that only tested existing approaches in case studies.
5 Results
The reviewed approaches vary substantially in inputs, encodings, architectures, prediction targets, and evaluation datasets. Recurrent networks are common, but the literature also includes CNNs, FFNNs, hybrid models, attention mechanisms, GANs, and multitask systems, making comparisons sensitive to design and data choices.
- Network architecture: Architectures are grouped into feed-forward, convolutional, and recurrent neural networks, with recurrent models generally preferred for temporally ordered activities.CNNs were also selected for simplicity and computational efficiency, while FFNNs require fixed-size inputs and therefore rescaling of prefixes.
- Data preprocessing & feature engineering: Input representations differ widely: studies select different process entities, encode variables mainly with one-hot encoding or embeddings, and use timestamps inconsistently.Most studies use activity information, while only a subset adds timestamps or other log attributes; some approaches also incorporate elapsed time, concurrent instances, or domain knowledge.
- Network architecture: Hybrid and specialized architectures include LRCN, attention-based models, differentiable neural computers, and GANs, with some reported improvements over prior or alternative models.Park and Song’s LRCN outperformed their LSTM and CNN alternatives, while attention combined separate activity and attribute representations before prediction.
- Data preprocessing & feature engineering: Fixed-size prefix handling can alter comparability: zero-padding is common, whereas dropping short prefixes caused more than half of the Helpdesk data to be excluded in one approach.The reported Helpdesk results used prefixes of size 4 and 6 despite a mean trace length of 3.6.
- Prediction target: Prediction targets mainly concern process classification and next-event, suffix, or time prediction, while execution cost and attribute-value estimation were not represented in the reviewed publications.Only a few studies performed multiple predictions, including multitask prediction of a time measure and the next event with one neural network.
- Evaluation: Evaluation spans 28 datasets, but most approaches use only a small subset; BPIC12 and Helpdesk are most frequent, limiting direct comparison across studies.Fourteen publications evaluated on only one or two process logs, while six studies used at least five datasets to examine robustness.
6 Discussion
The review finds substantial variation in inputs, architectures, evaluation practices, and dataset characteristics, limiting direct comparison of deep learning process-prediction approaches. Recurrent models dominate, but convolutional and hybrid architectures offer complementary capabilities.
- Input features: More input features generally improved performance, but only one publication used activity, timestamp, resource, and attribute information together.The review also identifies exceptions in which additional information reduced performance.
- Input features: Deep learning can incorporate sensory, textual, and image attributes, with pre-trained networks attachable through transfer learning.Text and image inputs generally require deeper networks to extract useful low-dimensional features.
- Input features: The review finds no definite encoding choice for categorical variables, with embedding benefits depending on vocabulary scale.Embedding advantages may be negligible for categories with 5 to 20 classes but may emerge at larger scales.
- Evaluation: Performance varies across logs, indicating that process-specific characteristics influence model behavior, while class imbalance can distort accuracy and favor dominant events.On Helpdesk, a naive model predicting only the final event reaches 37% accuracy, while some important activities occur below 1% frequency.
- Network architectures: Recurrent structures dominate process prediction because they model time dependency, although most studies train with fixed prefix sizes.The review notes that recurrent networks can theoretically process variable-length prefixes, while computational restrictions may motivate fixed sizes.
- Network architectures: Stacked LSTM architectures face vanishing-gradient difficulties, motivating supplementation with perceptrons or convolutional layers.LSTM tanh and sigmoid gates can contribute to vanishing gradients during backpropagation through deep networks.
- Network architectures: Convolutional and hybrid networks can learn elaborate process-log patterns, with one CNN reaching 93% accuracy on BPIC 12 and a hybrid outperforming its individual components.The cited CNN result is reported for BPIC 12, while the hybrid comparison concerns convolutional and recurrent layers.
- Evaluation: Quantitative comparisons are undermined by differing inputs, hyperparameter searches, test procedures, and metrics, especially accuracy on imbalanced datasets.The review calls for unified validation, benchmarks across public datasets, and additional metrics such as per-activity accuracy and temporal stability.
7 Future Challenges & Research Agenda
The paper identifies future research challenges spanning prediction-to-action, domain knowledge, input selection, and event representation. It emphasizes evaluating predictions in operational context and developing systematic ways to exploit richer process information.
- Prediction to action: Prediction is only a first step because process value depends on runtime adaptations that mitigate unwanted behavior.Possible adaptations include changing the next step, reallocating resources, or increasing execution priority.
- Prediction to action: Only two studies evaluated prediction quality through subsequent process adaptations, revealing a trade-off between prediction earliness and accuracy.Adaptation latencies make this trade-off practically relevant.
- Prediction to action: Future approaches should translate predictions into action suggestions because this holistic view may expose additional requirements.The paper notes that converting predictions into appropriate adaptations is non-trivial.
- Inputs and representations: Deep learning could exploit unstructured sensory, textual, and visual data, but broader inputs require systematic variable-selection methods.The explanatory power of attributes can vary across business cases and event logs.
- Inputs and representations: Current sequential event representations struggle with concurrent control-flow branches and loops, motivating revised event representations.The paper connects this limitation to the dependency implied when one event precedes another.
8 Conclusion
The review compares 32 deep learning approaches for process prediction and identifies strengths, weaknesses, and research gaps. Because datasets and evaluation methods rarely overlap, it provides a qualitative synthesis rather than a reliable performance ranking.
- 32 different deep learning approaches are compared using carefully selected criteria to identify strengths, weaknesses, and research gaps.
- The review classifies implementations by neural network type, prediction type, input features, and encoding methods.
- Sparse overlap in testing data and evaluation methods prevents a reliable quantitative ranking of process prediction approaches.A unified benchmark would enable quantitative comparison.
- The authors argue that broadening research across five major challenges could further improve existing approaches and support business process managers.