Tensor-Based Traffic Recovery of Industrial Network by BCD-Inspired Neural Approximation to Latent Nonlinearity and Sparsity

Due to the massive network data created by the industrial Internet of Things, tensor as the compact and multi-dimensional representation is widely employed to model the industrial network traffic. The unstable data acquisition often results in parts of entities' loss in the traffic tensor, which is conventionally solved by the tensor completion algorithms based on linear algebra and low rankness. However, due to the close correlation between domain transform and transformed sparsity, the linear domain transform cannot accurately approximate the latent nonlinear correlations which results in insufficient recovery performance. This article proposes a hybrid-structure deep model with nonlinear transform and sparse regularization to automatically search for the optimal domain transform approach and the corresponding sparse constraints. This model is in the tensor singular value decomposition framework and consists of two neural networks with different structures. One neural network has an autoencoder structure with fully connected networks to recover the lost entities only from partially observed data, and convolutional neural networks construct another one to constrain the sparsity in the transformed domain. Besides, we impose an additional Laplace constraint based on the local smoothness of the transformed tensor to overcome the continuous data loss. Inspired by the block coordinate descent algorithm, the mutually matching nonlinear transformer and sparse regularizers are trained alternatively. Extensive experimental results on industrial network traffic demonstrate that our proposed model outperforms the state-of-the-art approaches in different sampling rates and modes.