Missing traffic data imputation and pattern discovery with a Bayesian augmented tensor factorization model

Spatiotemporal traffic data, which represent multidimensional time series on considering different spatial locations, are ubiquitous in real-world transportation systems. However, the inevitable missing data problem makes data-driven intelligent transportation systems suffer from an incorrect response. Therefore, imputing missing values is of great importance but challenging as it is not easy to capture spatiotemporal traffic patterns, including explicit and latent features. In this study, we propose an augmented tensor factorization model by incorporating generic forms of domain knowledge from transportation systems. Specifically, we present a fully Bayesian framework for automatically learning parameters of this model using variational Bayes (VB). Relying on the publicly available urban traffic speed data set collected in Guangzhou, China, experiments on two types of missing data scenarios (i.e., random and non-random) demonstrate that the proposed Bayesian augmented tensor factorization (BATF) model achieves best imputation accuracies and outperforms the state-of-the-art baselines (e.g., Bayesian tensor factorization models). Besides, we discover interpretable patterns from the experimentally learned global parameter, biases, and latent factors that indeed conform to the dynamic of traffic states.