Active Few-shot Learning For RouteNet-Fermi

Xiaoyi Fu, Menghan Shao, Songze Li, Liuqing Yang

Abstract: Machine-learning-based Network Modeling requires a compact training data set that contains diversified network topology and configurations covering different congestion levels. We formalize the problem of modeling network delay using Multi-stage Message Passing Graph Neural Networks (GNNs) under the constraints of a limited number of training samples and a limited number of nodes for the topology of each sample as a few-shot learning problem. To tackle it, we propose an active learning algorithm that selectively randomizes initial features that are invariant of node numbers and then uses pool-based uncertainty sampling for selecting the approximated optimal network topology based on the Shannon entropy. The approximation could be theoretically proven and confirmed with experimental results.