---
title: "Remaining Cycle Time Prediction with Graph Neural Networks for Predictive Process Monitoring"
type: source
tags: [ppm, gnn, remaining-time, graph, deep-learning]
authors: [Duong Le Toan; Travé-Massuyès Louise; Subias Audine; Merle Christophe]
year: 2023
venue: "ICMLT 2023 (8th International Conference on Machine Learning Technologies), Stockholm"
kind: paper
raw_path: "raw/Predictive process monitoring/Graph neural nets 2023.pdf"
created: 2026-04-13
updated: 2026-04-13
key_claims:
  - Deep neural networks typically require data in Euclidean space; process logs can be represented as graphs instead.
  - Graph neural networks (GNN) learn relevant features automatically from this graph structure.
  - On real-life logs, GNN models improve RCT prediction accuracy over LSTM-based short-term memory state of the art, especially for complex processes.
---

# Duong et al. 2023 — GNN for Remaining Cycle Time Prediction

Applies **graph neural networks (GNN)** to [[concepts/remaining-time-prediction|remaining cycle-time prediction]]. Argues that traces are naturally graph-structured and that flattening them into sequences (for LSTM) loses structural information.

## Contribution
- GNN architecture operating directly on the graph representation of an event prefix.
- Outperforms LSTM short-term memory baselines, particularly on complex processes where the control flow branches heavily.

## Significance in the PPM landscape
Represents the **post-Transformer experimentation phase (2022+)** in PPM: after LSTMs (2017–2019) and Transformers (2020–2022), researchers began exploring GNNs and hybrid models seeking process-structure-aware inductive biases.

## Connections
**Concepts:** [[concepts/predictive-process-monitoring]] · [[concepts/remaining-time-prediction]] · [[concepts/lstm-ppm]] · [[concepts/transformer-ppm]]