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 M. Tavares, Thomas Seidl

In the world of Predictive Process Monitoring (PPM), the trend over the last decade has been clear: if you want to predict what happens next in a business process, throw a Deep Learning (DL) model at it. From LSTMs to Transformers and Reinforcement Learning, the complexity of these models has skyrocketed. But a new paper by Simon Rauch and his team at Ludwig-Maximilians-Universität München asks a provocative question: Do we actually need all that complexity?

Their answer is a framework called BEST (Bilaterally Expanding Subtrace Tree), and it suggests that structured, traditional data mining might actually be the superior choice for predicting the future of a process.

TL;DR

The researchers introduce BEST, a tree-based framework that predicts the remaining activities in a business process by mining and hierarchically structuring "bilaterally expanding" subtrace patterns. Despite having drastically lower model complexity than state-of-the-art LSTMs and Transformers, BEST achieves superior accuracy in remaining trace prediction and trains orders of magnitude faster.

What Problem Does This Paper Address?

Predictive Process Monitoring aims to look at a "running" case (like an insurance claim or a manufacturing sequence) and predict its future—specifically the Next Activity or the entire Activity Suffix (the remaining sequence until completion).

While Deep Learning models are powerful, they are "black boxes" that require massive computational resources (GPUs) and long training times. Furthermore, they can sometimes "hallucinate" transitions between activities that never actually occurred in the historical data. The authors set out to prove that a more transparent, pattern-based approach could be more efficient and more accurate by sticking strictly to sequences that have been observed in the real world.

Methods

The core of the approach is the Bilaterally Expanding Subtrace Pattern. Instead of just looking at what happened before an event (prefix), BEST looks at patterns that grow in both directions from a "center" activity.

  1. Bilateral Expansion: The algorithm starts with a single activity and "grows" the pattern by adding one activity to the left and one to the right. This captures the local context of an event more holistically.
  2. The BEST Tree: These patterns are organized into a hierarchical tree. The root is an empty pattern, level 1 contains single activities, and deeper levels contain increasingly longer, bilaterally expanded sequences.
  3. Bilateral Expansion Distance (BED): To navigate this tree, the authors developed a probabilistic distance metric. It combines the conditional probability of a pattern occurring with its size, allowing the model to find the "fittest" pattern to match a running case.
  4. Process Stages: To account for the fact that a process behaves differently at the start versus the end, they slice traces into "stages" and train specialized trees for different phases of the process.
  5. Pruning: To prevent the tree from exploding in size, they use a global probability cutoff. If a pattern is extremely rare, it’s pruned, keeping the model lightweight.

Datasets

The authors evaluated BEST against 11 real-world benchmark event logs, covering a wide range of domains:

  • Helpdesk: Support ticket logs.
  • BPI Challenges (2012 & 2013): Large-scale logs from financial and IT service management (including variations like 'Full', 'Work-flow only', and 'Closed Problems').
  • Env Permit: Environmental permit application processes.
  • Sepsis: Clinical pathways for hospital patients.
  • Nasa: Event logs from software testing.

Key Insights and Ideas

  • The "Ground Truth" Advantage: One reason BEST performs so well is that it only predicts sequences that have actually happened. DL models might predict a "statistically likely" next step that, in reality, is a forbidden transition in that specific business process.
  • Complexity vs. Performance: The authors found that by pruning patterns with a probability lower than 0.001%, they could reduce model complexity by 99% while maintaining steady performance.
  • Local vs. Global Context: For Next Activity Prediction (NAP), shorter patterns (smaller trees) actually work better because they focus on the immediate local context. For Remaining Trace Prediction (RTP), deeper trees are necessary to capture long-distance relationships between activities.
  • CPU Power: Unlike DL models that crave GPUs, BEST runs efficiently on standard CPUs, making it much more accessible for enterprise environments without specialized hardware.

Key Results

  • Superior Accuracy: In Remaining Trace Prediction (RTP), BEST outperformed state-of-the-art DL models (including LSTMs and Transformers like SuTraN) on the majority of the analyzed benchmarks.
  • Competitive NAP: For Next Activity Prediction, BEST was consistently in the top three performers, often beating established LSTM baselines.
  • Speedup: Training times for BEST are orders of magnitude faster than DL competitors. While a Transformer might take minutes or hours to train on a GPU, BEST often finishes in seconds on a CPU.
  • Inference Efficiency: Because the model is a structured tree, looking up a prediction is extremely fast, which is critical for real-time monitoring.

Limitations and Future Work

While BEST is a powerhouse for sequence prediction, it has a few "blind spots" the authors acknowledge:

  • Control-Flow Only: Currently, BEST only looks at the sequence of activities. It doesn't yet incorporate "data attributes" (like the cost of a step or the department performing it), which DL models can handle more easily.
  • Rigid Slicing: The "process stages" are currently sliced by fixed percentages of trace length. The authors suggest that "smarter" stage slicing based on actual process phases could improve results.
  • NAP Trade-off: As patterns get larger, the model's focus shifts away from the immediate next activity, leading to a slight drop in NAP accuracy. The authors propose investigating "asymmetrical" expansions (longer prefixes, shorter suffixes) to fix this.

In conclusion, BEST serves as a vital reminder for the data science community: sometimes, a well-structured tree is better than a deep forest of neurons.

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 M. 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