RFIN: A Recurrent Flow Interaction Network for Traffic Matrix Forecasting
Network traffic matrix forecasting estimates future network traffic demand from historical traffic observations, providing a basis for network resource allocation and operational management. A traffic matrix consists of multiple origin–destination (OD) flows, each representing the aggregated traffic between a pair of network nodes. Correlations between these flows can provide additional information for forecasting. An important question in traffic matrix forecasting is how to organize and use this information effectively. To address this question, we propose the recurrent flow interaction network (RFIN), which builds on flow-by-flow (FBF) temporal modeling and uses a small number of shared cores to exchange information among OD flows. This design avoids explicitly modeling a large number of pairwise relationships between flows. RFIN incorporates cross-flow aggregation and feedback into GRU recurrent state updates, enabling flow-level temporal modeling to use cross-flow context that evolves over time. We conducted experiments on two real-world traffic matrix datasets, Abilene and GÉANT, and compared RFIN with several representative forecasting methods. The results show that RFIN generally outperforms these methods in short-term traffic matrix forecasting, providing a practical approach to organizing and using cross-flow information.
Authors
- 叶东泉
- Minhua Shi
- Peifeng Qin
Institutions
- Innovation Academy for Microsatellites of Chinese Academy of Sciences (CN)
- Shanghai Ocean University (CN)
Publication Details
- Journal
- Electronics
- Published
- 2026-11-19
- DOI
- https://doi.org/10.3390/electronics15225393
- Primary Topic
- Traffic Prediction and Management Techniques
- Type
- article
- Field-Weighted Citation Impact
- 0.00