Abstract:

[Objective] This paper models the partially observed network performance data as a tensor, and exploits powerful ability of the deep neural network in feature extraction to recover the missing data. [Methods] Different from traditional tensor completion which relies on tensor decomposition, we design a novel tensor completion scheme based on Deep Convolution Autoencoder (DCAE). DCAE can handle the sparse matrix input, learn the complex relationship of data, and reconstruct the missing data. [Results] We have conducted extensive experiments using three publicly real-world network performance datasets. Our results demonstrate that DCAE can achieve significantly better recovery accuracy even when the sampling ratio is very low. [Limitations] Due to network attacks, network performance data may have unavoidable anomalies, which will deteriorate the recovery accuracy. In the future, we hope that the abnormal data will be able to be processed for a more robust recovery. [Conclusions] The proposed model can capture the non-linear relationship between the network performance data, achieve high data recovery accuracy, and recover missing data for advanced network applications.

Key words: tensor completion, sparse network monitoring, convolutional neural network, autoencoder, data recovery