Scalable Temporal Dimension Preserved Tensor Completion for Missing Traffic Data Imputation with Orthogonal Initialization

Abstract

Dear Editor, This letter puts forward a novel scalable temporal dimension preserved tensor completion model based on orthogonal initialization for missing traffic data (MTD) imputation. The MTD imputation acts directly on accessing the traffic state, and affects the traffic management. However, it still faces the following challenges: 1) The MTD imputation is usually formulated as matrix completion or tensor completion, which ignores the information across different dimensions; 2) Most of the existing models cannot generalize to traffic datasets of different scales or different missing rates; and 3) The MTD imputation models based on Gaussian random initialization easily leads to gradient explosion or vanishing, so that the training accuracy is not effectively improved. Inspired by these findings, the proposed scalable temporal dimension preserved tensor completion (ST-DPTC) model creatively establishes the following three-fold ideas: a) Incorporating the dimension preserved tensor completion (DPTC) to extract more distinctive traffic structure changes from the low-rank latent factor tensors; b) Adopting a scalable temporal (ST) regularization with first-order difference and second-order difference operators to adapt to different scales of traffic data; and c) Embedding ST regularization into DPTC with orthogonal initialization to perform low-rank latent factor tensor extraction and MTD imputation. Results on real-world traffic datasets with different scales show that our proposed model exceeds the state-of-the-art models in terms of the imputation accuracy.