A Multi-Attention Tensor Completion Network for Spatiotemporal Traffic Data Imputation

The widespread deployment of road sensors in the Internet of Things (IoT) allows for fine-grained data integration, which is a fundamental demand for data-driven applications. Sensing data with inevitable missing and substantial anomalies are unavoidable, due to unstable network communication, faulty sensors, etc. Recent tensor completion studies have demonstrated the superiority of deep learning in imputation tasks by precisely capturing the intricate spatiotemporal dependencies/correlations. However, ignoring the significance of initial interpolation in these methods results in unstable performance, especially for complicated missing scenarios across large-scale data. Additionally, the existing interpolation methods utilize recursive signal propagation along spatiotemporal dimensions, which produce noise accumulation where the dependencies are uncorrelated. In this study, we design a multiattention tensor completion network (MATCN) for modeling multidimensional representation in the presence of missing entries. MATCN sparsely sampled historical fragments and utilized a gated diffusion convolution layer to generate the initial schemes, which mitigate the exposure bias existing in previous traffic imputation models. In addition, we develop a spatial signal propagation module and a temporal self-attention module as the basic stack block of deep networks, which executes representation aggregation and dynamic dependencies extraction at the spatiotemporal level. This architecture empowers MATCN with progressive completion capacities for complex data missing scenarios. Numerical experiments on four real-world traffic data sets with various missing scenarios demonstrate the superiority of MATCN over multiple state-of-the-art imputation baselines.

Automated summary

In this study, a multiattention tensor completion network (MATCN) is designed to model multidimensional representation for handling missing data. MATCN generates initial schemes through sparse sampling of historical fragments and a gated diffusion convolution layer to address exposure bias in previous models. The architecture utilizes a spatial signal propagation module and a temporal self-attention module to aggregate representations and extract dynamic dependencies at the spatiotemporal level. Experimental results demonstrate the superiority of MATCN over other models in handling complex missing data scenarios.