Spatiotemporal subspace variational autoencoder with repair mechanism for traffic data imputation

High-quality spatial-temporal traffic data is crucial for the functioning of modern smart transportation systems. However, the collection and storage of traffic data in real-world scenarios are often hindered by many factors, causing data loss that greatly affects decision-making. Different modes of data absence result in varying degrees of information loss, which introduces considerable challenges to the precise imputation of traffic data. Many existing studies are concentrate on two main aspects: the examination of data distribution and the extraction of spatiotemporal relationships. On the one hand, methods that focus on distribution fitting do not require a large volume of observational data but often fail to capture spatial-temporal relationships, leading to overly smooth results. On the other hand, methods that aim to identify spatial-temporal relationships, while offering higher accuracy in fitting, demand a substantial amount of high-quality historical data. Taking into account the merits and demerits of both two paradigm, we developed a novel unsupervised two-stage model simultaneously takes into account the spatiotemporal distribution and relationships, termed Spatiotemporal Subspace Variational Autoencoder with Repair Mechanism (SVAE-R). In stage one, we introduced the concept of spatiotemporal subspace, which not only mitigates the noise impact caused by data sparsity but also reduces the cost for the model to find the distribution. In stage two, we designed a simple repair structure to capture spatial-temporal relationships among data through graph convolution network(GCN) and gated recurrent units(GRU), revising the details of the data. We have evaluated our model on two authentic datasets, and it has exhibited a high degree of robustness, maintaining effective performance even under extreme data loss conditions.