Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

PGTNet: A Process Graph Transformer Network for Remaining Time Prediction of Business Process Instances

Keyvan Amiri Elyasi, Han van der Aa, Heiner Stuckenschmidt

TL;DR

PGTNet addresses remaining time prediction in predictive process monitoring by transforming event logs into graphs of event prefixes and training a GPS Graph Transformer to perform graph regression with $L_1$ loss. The method endows nodes with event-class semantics and edges with directly-follows relations plus rich temporal and attribute features, enabling simultaneous learning of local control-flow and long-range dependencies. Across 20 real-world logs, PGTNet achieves a substantial accuracy boost (average $MAE$ around $12.92$ days) compared with state-of-the-art baselines, with especially large gains on highly complex processes. The approach supports multi-perspective data fusion, shows promising earliness in predictions, and offers reproducibility via a public repository, underscoring its practical impact for predictive process monitoring.

Abstract

We present PGTNet, an approach that transforms event logs into graph datasets and leverages graph-oriented data for training Process Graph Transformer Networks to predict the remaining time of business process instances. PGTNet consistently outperforms state-of-the-art deep learning approaches across a diverse range of 20 publicly available real-world event logs. Notably, our approach is most promising for highly complex processes, where existing deep learning approaches encounter difficulties stemming from their limited ability to learn control-flow relationships among process activities and capture long-range dependencies. PGTNet addresses these challenges, while also being able to consider multiple process perspectives during the learning process.

PGTNet: A Process Graph Transformer Network for Remaining Time Prediction of Business Process Instances

TL;DR

PGTNet addresses remaining time prediction in predictive process monitoring by transforming event logs into graphs of event prefixes and training a GPS Graph Transformer to perform graph regression with loss. The method endows nodes with event-class semantics and edges with directly-follows relations plus rich temporal and attribute features, enabling simultaneous learning of local control-flow and long-range dependencies. Across 20 real-world logs, PGTNet achieves a substantial accuracy boost (average around days) compared with state-of-the-art baselines, with especially large gains on highly complex processes. The approach supports multi-perspective data fusion, shows promising earliness in predictions, and offers reproducibility via a public repository, underscoring its practical impact for predictive process monitoring.

Abstract

We present PGTNet, an approach that transforms event logs into graph datasets and leverages graph-oriented data for training Process Graph Transformer Networks to predict the remaining time of business process instances. PGTNet consistently outperforms state-of-the-art deep learning approaches across a diverse range of 20 publicly available real-world event logs. Notably, our approach is most promising for highly complex processes, where existing deep learning approaches encounter difficulties stemming from their limited ability to learn control-flow relationships among process activities and capture long-range dependencies. PGTNet addresses these challenges, while also being able to consider multiple process perspectives during the learning process.
Paper Structure (10 sections, 4 equations, 4 figures, 3 tables)

This paper contains 10 sections, 4 equations, 4 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Background and Related work
  3. Preliminaries
  4. PGTNet for Remaining Time Prediction
  5. Graph Representation of Event Prefixes
  6. Training PGTNet to Predict Remaining Time
  7. Evaluation
  8. Experimental Setup
  9. Results
  10. Conclusion and Future Work

Figures (4)

  • Figure 1: Graph representation of an event prefix of length of 6, Case ID= '27583'.
  • Figure 2: PGTNet architecture: based on the GPS Graph Transformer recipe rampasek_recipe_2022. Paths to process node and edge features are specified by blue and red colors, respectively.
  • Figure 3: Remaining time prediction accuracy in terms of relative MAE (in percentage).
  • Figure 4: MAE over different prefix lengths (selected event logs).