Network traffic matrix prediction with incomplete data via masked matrix modeling

Traffic matrix (TM) prediction aims to forecast future traffic data for networks using historical traffic matrices. TM prediction plays a critical role in solving network engineering tasks such as routing management, capacity planning, and network security. Previous research assumes that the TM sequences fed into the prediction models are complete and precise. However, considering the unavoidable failure of network systems and their large monitoring costs, it is impractical to collect complete TMs from large-scale networks. In this paper, we study the TM prediction problem with randomly missing values. To perform TM completion, we introduce a masked matrix modeling method based on self-supervised learning that can learn better matrix representations. As the matrix completion and prediction tasks are highly correlated, we propose an end-to-end framework that performs the two tasks simultaneously through joint learning. Specifically, we design a 3D-UNet architecture that is able to exploit multi-scale spatio–temporal correlations in a TM sequence as the completion module. An LSTM2D architecture is employed as the prediction module to take advantage of spatio–temporal dependencies. Extensive experiments are conducted on publicly available datasets, and the results show that our model significantly outperforms previous state-of-the-art methods. Source code is available at https://github.com/FreeeBird/TM_prediction_random_missing.