A Distributed Regularization Method for Latency Matrix Completion in the Wild

Latency matrix completion is an important foundation of latency-sensitive applications optimization. On the basis of the in-depth discussion of the robustness of a kind of matrix-factorization based non-gradient descending completion methods, this paper analyzes the significant impact to the intrinsic ill-posed and ill-conditioned inverse problems in the methods caused by the oscillations of the latency sequences. To mitigate the impact and improve the performance of the matrix completion methods in the wild, a regularization factor is introduced to improve the spectrum signature of the coefficient matrix, a median-Kalman filter, a time-spatial federated filtering scheme, is proposed to smooth the latency sequences, and then the topology mutation is obtained through extracting the statistic characters of the latency sequences. The experiments show that our method can avoid the performance degradation caused by noises without losing the major characteristics of the latency sequences, provide robust latency estimation capability, and keep the stress coefficient at a low level about 0.13 during the whole life cycle of the network.