Hierarchical Structuring of Bilaterally Expanding Subtrace Patterns for Efficient Tree-based Activity Suffix Prediction

Simon Rauch, Christian M. M. Frey, Andrea Maldonado, Daniel Schuster, Gabriel Marques Tavares, Thomas Seidl

Abstract

Business processes of different domains are potentially prone to high variability, leading to high amounts of uncertainty with respect to future case execution. The field of predictive process monitoring has recognized this fate and developed handling this uncertainty as one its core concerns for building reliable predictive or prescriptive methods. Over the last decade, deep learning methods have increasingly emerged as a product of this research field and are considered as the preferred approach when it comes to the prediction of next activities or activity suffixes. However, it remains an open question whether deep learning models finally surpass traditional data mining techniques for these tasks. We address this question in our paper by proposing a framework for process event sequence prediction framework which is based on a hierarchical structuring of bilaterally expanding subtraces mined from activity traces which takes their structural relationship and inter-pattern distances into account. The resulting tree structure serves as an efficient alternative approach for currently established deep learning techniques due to its drastically lower model complexity. The hierarchical arrangement can directly be leveraged for forecasting the most probable future activities given the recent trace history. We achieve competitive forecasting results for remaining trace prediction, even surpassing state-of-the-art deep learning approaches on the majority of the analyzed real-world benchmark process event logs while only relying on the available control-flow information.

Introduction

Predictive Process Monitoring (PPM) aims to forecast the future behavior of ongoing business processes, such as predicting the next activity or the entire remaining activity suffix. Recently, the field has been dominated by complex Deep Learning (DL) architectures, including LSTMs and Transformers. While powerful, these models often act as "black boxes," require significant computational resources, and can occasionally predict transitions that never occurred in historical data. This research introduces the Bilaterally Expanding Subtrace Tree (BEST) framework, a pattern-based approach that leverages structured data mining to provide a more efficient and transparent alternative to DL models.

Methodology: The BEST Framework

The BEST framework is built upon the concept of bilaterally expanding subtrace patterns. Unlike traditional methods that focus solely on prefixes, BEST analyzes patterns that grow in both directions from a central activity, capturing a more holistic local context.

Schematic overview of the BEST framework, illustrating the hierarchical structuring of bilaterally expanding subtrace patterns.

Core Components

  1. Bilateral Expansion: Patterns are initialized with a single activity and expanded by adding activities to both the left and right. This captures the surrounding context of an event more effectively than unidirectional prefixes.
  2. Hierarchical Tree Structure: These patterns are organized into a tree where the root is an empty pattern, and subsequent levels contain increasingly longer, expanded sequences.
  3. Bilateral Expansion Distance (BED): A probabilistic distance metric is used to navigate the tree. It combines the conditional probability of a pattern's occurrence with its size to identify the most relevant pattern for a given running case.
  4. Process Stages: To account for temporal variations in process behavior, traces are sliced into stages, and specialized trees are trained for different phases of the process lifecycle.
  5. Pruning: To maintain efficiency, a global probability cutoff is applied. Patterns with extremely low frequency are pruned, reducing model complexity by up to 99% without significant performance loss.

Experimental Evaluation

The framework was evaluated using 11 real-world benchmark event logs from diverse domains, including IT service management (BPI Challenges), healthcare (Sepsis), and software testing (Nasa). BEST was compared against state-of-the-art DL baselines, including LSTMs and specialized Transformers like SuTraN.

Results and Discussion

The experimental results demonstrate that BEST is a highly competitive alternative to deep learning for process prediction:

  • Accuracy: In Remaining Trace Prediction (RTP), BEST outperformed state-of-the-art DL models on the majority of the analyzed benchmarks. For Next Activity Prediction (NAP), it consistently ranked among the top performers.
  • Efficiency: Training BEST is orders of magnitude faster than DL models. While Transformers may require minutes or hours on a GPU, BEST typically completes training in seconds on a standard CPU.
  • Reliability: By sticking strictly to observed historical sequences, BEST avoids the "hallucination" of impossible transitions, a common issue with probabilistic DL models.

Conclusion

BEST serves as a vital reminder that well-structured, traditional data mining techniques can match or exceed the performance of complex deep learning models in specific domains. By prioritizing observed patterns and hierarchical organization, BEST offers a transparent, efficient, and highly accurate solution for real-time predictive process monitoring. Future work will explore the integration of data attributes beyond control-flow information and more dynamic methods for process stage slicing.

Summary updated

Abstract

<title>Abstract</title> Business processes of different domains are potentially prone to high variability, leading to high amounts of uncertainty with respect to future case execution. The field of predictive process monitoring has recognized this fate and developed handling this uncertainty as one its core concerns for building reliable predictive or prescriptive methods. Over the last decade, deep learning methods have increasingly emerged as a product of this research field and are considered as the preferred approach when it comes to the prediction of next activities or activity suffixes. However, it remains an open question whether deep learning models finally surpass traditional data mining techniques for these tasks. We address this question in our paper by proposing a framework for process event sequence prediction framework which is based on a hierarchical structuring of bilaterally expanding subtraces mined from activity traces which takes their structural relationship and inter-pattern distances into account. The resulting tree structure serves as an effiecient alternative approach for currently established deep learning techniques due to its drastically lower model complexity. The hierarchical arrangement can directly be leveraged for forecasting the most probable future activities given the recent trace history. We achieve competitive forecasting results for remaining trace prediction, even surpassing state-of-the-art deep learning approaches on the majority of the analyzed real-world benchmark process event logs while only relying on the available control-flow information.
BibTeX
@article{rauch2026hierarchical,
  author = {Simon Rauch and Christian M. M. Frey and Andrea Maldonado and Daniel Schuster and Gabriel Marques Tavares and Thomas Seidl},
  title = {{Hierarchical Structuring of Bilaterally Expanding Subtrace Patterns for Efficient Tree-based Activity Suffix Prediction}},
  journal = {Process Sci.},
  year = {2026},
  doi = {10.1007/s44311-026-00050-y},
}
Figure