Multi-View Tailored Tensor Completion for Spatiotemporal Traffic Data Imputation

Paper in proceeding, 2025

Advanced sensing technologies have enabled multiview observations of traffic dynamics, yet data missingness remains a significant challenge to reliable traffic monitoring. Unlike existing single-view imputation approaches, we propose Multi-View Tailored Tensor Completion (MVT2C), a framework integrating intra-view and inter-view modules. The intra-view module employs third-order tensor schemes to characterize highdimensional low-rank properties within each view, while the inter-view module captures view-to-view relationships through subspace representation and accommodates view-specific discrepancies via structured column-sparse reconstruction error matrices. We formulate the integrated framework as a multiblock non-convex optimization problem, solved efficiently using an inexact augmented Lagrangian multiplier method. Experiments on two real-world datasets demonstrate the superiority of MVT2C across various missing scenarios compared to state-of-the-art baselines. Further analyses reveal that leveraging crossview information not only improves robustness and stability but also yields affinity matrices that reflect both shared and unique view-specific features.