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.
Introduction
In business process management, accurately predicting the remaining time of a case—such as a loan application or a medical patient's journey—is vital for resource planning and deadline management. While deep learning has advanced this field, existing sequence-based models like LSTMs and standard Transformers often struggle with the non-linear nature of business processes, which frequently involve loops, parallel tasks, and complex control-flow logic. LSTMs are limited by long sequences, while standard Transformers may ignore explicit process structures. Conversely, standard Graph Neural Networks (GNNs) can capture structure but often suffer from "over-smoothing," losing detail over multiple steps.
Methodology
PGTNet (Process Graph Transformer Network) addresses these limitations through a two-step hybrid approach that combines structural awareness with long-range dependency tracking.
Graph Transformation
Instead of treating process data as simple sequences, PGTNet transforms process "prefixes" (the history of a running case) into directed attributed graphs. In this representation, nodes correspond to unique activities, and edges represent "directly-follows" relationships. A key innovation is the use of rich edge features, which incorporate temporal data (e.g., duration, time of day), case attributes (e.g., claim amounts), and system workload metrics (active cases).
Hybrid Architecture
The architecture is based on the GPS (General, Powerful, Scalable) Graph Transformer recipe. Each layer consists of two parallel blocks designed to capture different process dynamics:
- MPNN (Message Passing Neural Network): Utilizes a Graph Isomorphism Network (GIN) to capture local control-flow relationships.
- Global Attention (Transformer): Enables every node to interact with every other node, capturing long-range dependencies regardless of the graph's physical structure.
- Positional Encodings: Laplacian eigenvectors and random-walk structural encodings help the model understand the "shape" and relative position of activities within the process graph.
Evaluation and Key Results
PGTNet was evaluated on 20 publicly available real-world event logs, including BPIC logs (banking, environmental permits), medical logs (Sepsis), and municipal regulation logs (Traffic Fines).
- Superior Accuracy: PGTNet achieved an average Mean Absolute Error (MAE) of 12.92 days, a significant improvement over the next best approach (GGNN) at 24.63 days. This represents a roughly 47% reduction in error.
- Performance in Complexity: The model shows a strong "complexity synergy," performing exceptionally well on "messy" processes with high entropy where standard models fail to capture sparse but vital relationships.
- Earliness and Efficiency: PGTNet provides reliable estimates much earlier in the process life cycle than baseline models. Despite its complexity, it maintains a fast inference time of approximately 3 milliseconds per prediction.
Discussion and Future Work
Ablation studies highlight that while additional data attributes (edge-centric encoding) improve performance, the hybrid GNN-Transformer architecture itself is the primary driver of accuracy gains. By moving case-specific attributes to edges, the model maintains consistent node definitions across different cases, preventing graph sparsity.
While PGTNet requires more training time than simple LSTMs (averaging 12.9 hours vs. 1.6 hours), its predictive power makes it a compelling choice for complex environments. Future research will explore extending this graph-based representation to next-activity prediction and adapting the model for Object-Centric Event Logs (OCEL) to track multiple interacting objects simultaneously.
Summary updated