Vehicular Crowdsensing Inference and Prediction With Multi Training Graph Transformer Networks

Vehicular crowdsensing has emerged as a prominent sensing paradigm in the Internet of Things (IoT), and its powerful sensing and computing capabilities can provide sufficient data for various applications. To reduce the cost while ensuring the sensing quality, sparse mobile crowdsensing has been proposed, which only requires data from some sensing areas and utilizes spatiotemporal correlation to infer data for other unsensed areas. In real vehicular crowdsensing scenarios, not only the current period sensing data is required to be inferred but also the prediction of the future whole sensing map is of great significance. In this article, we propose multitask pretraining graph transformer networks (MT-PTGTN) that incorporate graph neural networks (GNNs) and transformer to support both data inference and prediction for vehicular crowdsensing. Comprehensively considering the surface and underlying patterns among the sensing grids, MT-PTGTN utilizes pre-training topological mining GNNs and graph attention networks to model two patterns, respectively, which contributes to improving inference accuracy. The transformer with the multilayer attention mechanism is incorporated to capture the temporal correlation of the sensing data and predict the future complete sensing map. Furthermore, to prevent the error propagation from the inference task to the subsequent prediction task, we propose a dynamic multi-task learning framework that dynamically adjusts the weights of tasks during training. The experimental evaluation on the real-world dataset demonstrates the superiority of MT-PTGTN in data inference and prediction.