Flow Neural Network and Flow-Structured Data Representation
Traffic flows are the most fundamental components in a communication networking system. An accurate understanding of these flows is crucial for many downstream network applications. However, the high nonlinearity, randomness and complicated self similarity (Leland et al., 1994) of these traffic thwart extensive traditional analytical and learning models, particularly in high time resolution. In this paper, we experimentally discover the spatio-temporal induction effect, a universally present property in flow-structured network traffic data. By exploiting the flowing invariance and variance, we propose FlowNN and induction operation to learn the representations of flow-structured data in communication networks. Results from regression tasks demonstrate several times better accuracy than the state-of-the-art baselines and extend the existing time resolution from second to millisecond, offering a good testament to the strength of our approach.
PDF Abstract