Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Time-Aware and Transition-Semantic Graph Neural Networks for Interpretable Predictive Business Process Monitoring

Fang Wang, Ernesto Damiani

TL;DR

The paper tackles predictive business process monitoring by unifying prefix-based GCNs and full-sequence GATs within a time-aware, semantically enriched graph framework. It introduces a time-decay attention mechanism to form dynamic, prediction-centered windows and embeds transition semantics into edge features, enhancing both accuracy and interpretability. Empirical results across six PBPM datasets show that global attention with temporal and semantic cues yields competitive performance and that prefix-based GCNs remain strong in data-sparse, high-cardinality regimes, offering a practical guideline for model selection. The work also delivers multilevel interpretability tools, including attention heatmaps and adaptive window analyses, making the framework actionable for process optimization and compliance evaluation. Overall, the approach advances PBPM by balancing architectural exploration, temporal dynamics, semantic meaning, and transparent explanations in a scalable, generalizable manner.

Abstract

Predictive Business Process Monitoring (PBPM) aims to forecast future events in ongoing cases based on historical event logs. While Graph Neural Networks (GNNs) are well suited to capture structural dependencies in process data, existing GNN-based PBPM models remain underdeveloped. Most rely either on short prefix subgraphs or global architectures that overlook temporal relevance and transition semantics. We propose a unified, interpretable GNN framework that advances the state of the art along three key axes. First, we compare prefix-based Graph Convolutional Networks(GCNs) and full trace Graph Attention Networks(GATs) to quantify the performance gap between localized and global modeling. Second, we introduce a novel time decay attention mechanism that constructs dynamic, prediction-centered windows, emphasizing temporally relevant history and suppressing noise. Third, we embed transition type semantics into edge features to enable fine grained reasoning over structurally ambiguous traces. Our architecture includes multilevel interpretability modules, offering diverse visualizations of attention behavior. Evaluated on five benchmarks, the proposed models achieve competitive Top-k accuracy and DL scores without per-dataset tuning. By addressing architectural, temporal, and semantic gaps, this work presents a robust, generalizable, and explainable solution for next event prediction in PBPM.

Time-Aware and Transition-Semantic Graph Neural Networks for Interpretable Predictive Business Process Monitoring

TL;DR

The paper tackles predictive business process monitoring by unifying prefix-based GCNs and full-sequence GATs within a time-aware, semantically enriched graph framework. It introduces a time-decay attention mechanism to form dynamic, prediction-centered windows and embeds transition semantics into edge features, enhancing both accuracy and interpretability. Empirical results across six PBPM datasets show that global attention with temporal and semantic cues yields competitive performance and that prefix-based GCNs remain strong in data-sparse, high-cardinality regimes, offering a practical guideline for model selection. The work also delivers multilevel interpretability tools, including attention heatmaps and adaptive window analyses, making the framework actionable for process optimization and compliance evaluation. Overall, the approach advances PBPM by balancing architectural exploration, temporal dynamics, semantic meaning, and transparent explanations in a scalable, generalizable manner.

Abstract

Predictive Business Process Monitoring (PBPM) aims to forecast future events in ongoing cases based on historical event logs. While Graph Neural Networks (GNNs) are well suited to capture structural dependencies in process data, existing GNN-based PBPM models remain underdeveloped. Most rely either on short prefix subgraphs or global architectures that overlook temporal relevance and transition semantics. We propose a unified, interpretable GNN framework that advances the state of the art along three key axes. First, we compare prefix-based Graph Convolutional Networks(GCNs) and full trace Graph Attention Networks(GATs) to quantify the performance gap between localized and global modeling. Second, we introduce a novel time decay attention mechanism that constructs dynamic, prediction-centered windows, emphasizing temporally relevant history and suppressing noise. Third, we embed transition type semantics into edge features to enable fine grained reasoning over structurally ambiguous traces. Our architecture includes multilevel interpretability modules, offering diverse visualizations of attention behavior. Evaluated on five benchmarks, the proposed models achieve competitive Top-k accuracy and DL scores without per-dataset tuning. By addressing architectural, temporal, and semantic gaps, this work presents a robust, generalizable, and explainable solution for next event prediction in PBPM.

Paper Structure

This paper contains 52 sections, 12 equations, 20 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. From Prefix to Global Models
  4. Temporal Dynamics in Process Prediction
  5. Interpretability in PBPM
  6. Preliminaries and Definitions
  7. Research Pipeline and Methodology
  8. Modeling Pipeline Overview
  9. Graph Convolutional Networks with Prefix Subgraphs
  10. Graph Convolutional Networks
  11. Baseline Prefix GCN Model
  12. Input and Output Structure
  13. Model Architecture
  14. Summary
  15. Graph Attention Networks with Full Sequence Graphs
  16. ...and 37 more sections

Figures (20)

  • Figure 1: Overview of the proposed framework. Event logs are transformed into attributed event-sequence graphs capturing local time gaps ($\Delta t$), global time decay ($\boldsymbol{\delta}$), and transition semantics ($\tau$). Prefix-based GCNs model local temporal dependencies across multiple prefix sizes, while full-trace GATs progressively incorporate temporal and semantic cues through layered attention mechanisms. The outputs correspond to next-event predictions for GCNs and sequence of next-event generation for GATs, with interpretability provided in the dynamic temporal (decay) based variants.
  • Figure 2: Variants arranged by temporal cue (rows: local $\Delta t$ vs global decay $\delta$) and transition semantics (columns: without vs with $\tau$). Arrows denote incremental extensions: GAT-T $\xrightarrow{+\tau}$ GAT-TT; GAT-T $\xrightarrow{\Delta t \to \delta}$ GAT-TD; GAT-TD $\xrightarrow{+\tau}$ GAT-TDTE (final).
  • Figure 3: Baseline GCN Models Performance with Different Prefix lengths
  • Figure 4: Interpretability Analysis of the GAT-TD Model on the BPI13i Dataset
  • Figure 5: Correlation between Edge type Scores and Attention Weights in the GAT-TDTE model on BPI13i Dataset
  • ...and 15 more figures

Theorems & Definitions (8)

  • Definition 1: Event-Sequence Graph Construction
  • Definition 2: Event Label Vector
  • Definition 3: Node Attribute Vector
  • Definition 4: Edge Attributes Vector
  • Definition 5: Edge Index
  • Definition 6: Global Temporal Distance Vector
  • Definition 7: Graph Attribute Vector
  • Definition 8: Prefix Subgraph