Conditional Invertible Neural Networks for Traffic Matrix Estimation from Link Loads

Traffic matrices (TMs) are essential to many network management and optimization tasks. However, direct measurement of traffic flows is impractical, and a more viable alternative is traffic matrix estimation (TME) via network tomography, inferring TMs from readily available link load measurements and routing information. In this paper, we address this ill-posed linear inverse problem by leveraging conditional invertible neural networks (cINNs), a recent neural architecture that introduces conditioning into its building blocks to approximate the posterior conditional distribution of TMs given the observed link loads. Specifically, we explore two alternative conditioning mechanisms: directly integrating raw link loads or incorporating extracted features generated by a dedicated conditioning neural network. The trained cINN enables efficient sampling from the posterior conditional distribution, allowing accurate TM estimation via either computing the conditional mean or iteratively solving a minimization problem to ensure data fidelity with the link measurements. The proposed TME approach is evaluated on a dataset obtained from a real-world backbone network and compared against three baseline methods across a comprehensive set of performance metrics. A prototype implementation is publicly available under an open-source license.