Dual-domain low-rank tensor completion for traffic data recovery

Complete traffic data plays an indispensable role in intelligent transportation network management that has great potential to improve traffic safety and alleviate road congestion. However, due to the malfunctions in sensing devices or communication networks, missing data phenomena are ubiquitous in the real world and thus pose formidable challenges to network management. Recent studies successfully exploit a low-rank property of traffic tensor data for missing data recovery. However, they focus on either original domain or transformed domain, which may not suffice to exploit abundant tensor structure information. To contend with this problem, this article proposes a novel dual-domain nonconvex low-rank tensor completion to simultaneously take advantage of the low-rankness of traffic data tensor in both original domain and transformed domain. Specifically, we first devise two nonconvex tensor norms that not only characterize the global low-rank property in both domains but also capture the multi-dimensional correlation along different modes. Then, we further incorporate the local temporal consistency as regularization to leverage the intrinsic temporal continuity. By doing so, our model comprehensively leverages the global low-rankness and local temporal property. To solve the proposed model, following the alternating direction method of multipliers framework, we develop an efficient iterative algorithm and analyze the computational complexity. Numerous experiments on real-world datasets under diverse missing patterns and missing ratios verify the superiority of our algorithm over some leading baseline methods.