Forecasting Individual NetFlows using a Predictive Masked Graph Autoencoder
Forecasting Individual NetFlows using a Predictive Masked Graph Autoencoder
Authors: Georgios Anyfantis, Pere Barlet-Ros Date: 2026-04-22 Paper ID: openalex:2604.20483
Summary
This paper proposes a Graph Neural Network (GNN) approach for predicting individual network flow (NetFlow) traffic by modeling network activity as a series of heterogeneous bidirectional graphs. The model uses sliding windows to capture temporal dependencies between IP, Port, and Connection nodes, enabling both structural and feature-based forecasting. Results demonstrate that the GNN framework excels at predicting connection attachment points while maintaining competitive performance on feature reconstruction tasks compared to established forecasting baselines.
Key Contributions
- Introduces a GNN-based framework for per-flow NetFlow traffic prediction using heterogeneous graph representations of network connections.
- Demonstrates effective modeling of temporal evolution of both graph topology (IP/Port connections) and connection-specific features.
- Achieves superior performance in predicting flow attachment points (IP/Port) compared to standard time-series baselines.
Open Questions & Future Work
Archivist Review
The paper provides a specific application of GNNs to NetFlow prediction, but the architecture itself is a standard composition of known techniques (GNNs + masked autoencoders) rather than a novel, reusable concept. The open question regarding variable-size graph forecasting is a meaningful bottleneck in spatiotemporal network modeling, thus it was approved.
Approved Open Questions
- Variable Graph Size Forecasting: Fixed-size constraints limit the scalability and realism of graph-based network traffic forecasting, acting as a critical bottleneck for production-grade deployment.
Rejected Candidates
- [concept] Predictive Masked Graph Autoencoder (
predictive-masked-graph-autoencoder) - not_novel: The proposed architecture is a straightforward application of masked autoencoders to graphs, lacking the distinct conceptual novelty required for a standalone vault entry.
Links
Metadata & Links
- url
- https://arxiv.org/abs/2604.20483
- paper_id
- 2604.20483
- paper_source
- openalex
- domain
- nlp
- tags
-
- architectures
-
- datasets
-
- concept_slugs
-
- dataset_slugs
-
- skill
- TimeSeriesSkill
- created_at
- 2026-04-25T05:37:46Z