Abstract
Traffic Speed Prediction (TSP) is decisive factor for Intelligent Transportation Systems (ITS), targeting to estimate the traffic speed depending on real-time data. It enables efficient traffic management, congestion reduction, and improved urban mobility in ITS. However, some of the challenges of TSP are dynamic nature of temporal and spatial factors, less generalization, unstable and increased prediction horizon. Among these challenges, the traffic speed prediction is highly challenged due to complicated spatiotemporal dependencies in road networks. In this research, a novel approach called Multi Objective Graph Learning (MOGL) includes the Adaptive Graph Sampling with Spatio Temporal Graph Neural Network (AGS-STGNN), Pareto Efficient Global Optimization (ParEGO) as multi objective Bayesian optimization in adaptive graph sampling and enhanced Attention Gated Recurrent Units (EAGRU). The proposed MOGL approach is composed of three phases. The first phase is an AGS-STGNN for selecting the optimal samples of both spatial and temporal. Second phase has a ParEGO-based adaptive graph sampling is employed to extract spatial features that are refined through GNN layers by selecting temporal features. Third phase is an EAGRU module to enhance feature representation and prediction reliability by incorporating the feature fusion stage to dynamically prioritizing the critical road segments and time intervals, allowing the model to focus on the most influential factors affecting traffic speed variations in prediction. The supremacy of the proposed model is validated using the two available benchmark traffic datasets METR-LA and PeMS-BAY. The experimental results of the proposed MOGL approach is evaluated in terms accuracy, Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Mean Absolute Error (MAE) to predict and estimate the traffic speed. It ensures the proposed approach offers a better real-time and large-scale traffic speed prediction compared to existing approaches with the values 2.09 and 2.15 for MAE, 3.29 and 3.22 for RMSE and 3.17 and 3.21 for MAPE in both datasets. The proposed approach showed higher performance, notably on the METR-LA dataset, with a significant RMSE reduction of up to 28.9% compared to DSTMAN, outperforming other models like STGCN variants.