Street_Car_1297 seeks community input on designing a GNN solution for next activity prediction in business process mining, sharing project structure, dataset specifics, and open modeling questions.

Structuring a GNN-Based Next Activity Prediction for Business Process Mining

Posted by Street_Car_1297

This post discusses the early stages of a project aimed at using Graph Neural Networks (GNNs) to predict the next event in business process data, specifically helpdesk support traces stored in CSV files.

Project and Data Overview

  • Goal: Build a model for next activity prediction at the node level using graph representations of business processes.
  • Input Data: A CSV representing process traces—each trace becomes a graph (a Directly-Follows Graph) where nodes represent events or activities.
  • Dataset Size: 4,580 separate process instances (graphs) with an average of 7 nodes each and 15 total activity types (labels).

Core Modeling Questions

  1. Model Architecture: Considering a 3-layer Graph Convolutional Network (GCN), but open to community suggestions for architectures better suited to sequence-based node prediction within process graphs.
  2. Handling Multiple Instances: Unsure whether to treat 4,580 traces as separate graphs for training or to merge them into a single large graph while preserving per-node instance information. Also seeking advice on batching strategies for GNNs with many small graphs versus constructing a global graph for training.

Additional Context

  • The use case involves learning to predict the next activity in what is essentially a sequence, but modeled as a graph structure to capture more complex dependencies than standard sequence models.
  • The author notes a lack of practical examples or published studies directly matching this scenario, highlighting both the novelty and challenge of the task.

Community Collaboration

  • Street_Car_1297 encourages feedback, especially from those experienced in GNNs or process data, looking for recommendations on model structure, data batching, and best practices for graph-based activity prediction.

Key Takeaways:

  • Focused on the technical modeling of business process data using GNNs
  • Explicitly mentions challenges around architecture and multi-instance data handling
  • Targeted at practitioners interested in advanced machine learning techniques for process mining

This post appeared first on “Reddit Machine Learning”. Read the entire article here