Semantic understanding and prompt engineering for large-scale traffic data imputation

Intelligent Transportation Systems (ITS) face the formidable challenge of large-scale missing data, particularly in the imputation of traffic data. Existing studies have mainly relied on modeling network-level spatiotemporal correlations to address this issue. However, these methods often overlook the rich semantic information (e.g., road infrastructure, sensor location, etc.) inherent in road networks when capturing network-wide spatiotemporal correlations. We address this limitation by presenting the Graph Transformer-based Traffic Data Imputation (GT-TDI) model, which imputes missing values in extensive traffic data by leveraging spatiotemporal semantic understanding of road networks. The proposed model leverages semantic descriptions that capture the spatial and temporal dynamics of traffic across road networks, enhancing its capacity to infer comprehensive spatiotemporal relationships. Moreover, to augment the model's capabilities, we employ a Large Language Model (LLM) and prompt engineering to enable natural and intuitive interactions with the traffic data imputation system, allowing users to query and request in plain language, without requiring expert knowledge or complex mathematical models. The proposed model, GT-TDI, utilizes Graph Neural Networks (GNN) and Transformer architectures to perform large-scale traffic data imputation using deficient observations, sensor social connectivity, and semantic descriptions as inputs. We evaluate the GT-TDI model on the PeMS freeway dataset and benchmark it against cutting-edge models. The experimental evidence demonstrates that GT-TDI surpasses the cutting-edge approaches in scenarios with intricate patterns and varying rates of missing data.