An accepted paper published in Scientific Reports on 10 October by researchers affiliated with King Abdulaziz University and Taif University proposes a graph-attention network combining spatial and temporal relationships to classify traffic congestion. Called DLUMIS-TCF, the system uses Elastic Net for feature selection, a heterogeneous spatiotemporal graph-attention network for classification, AdaBelief for optimization and SHAP to explain feature contributions. Evaluation used the public Urban Mobility Intelligence dataset from Kaggle, containing 3,287 trip-level records. The paper does not specify a city for this dataset.

The paper reports 94.07% accuracy on an 80/20 test split and 94.15% mean accuracy in ten-fold cross-validation. These results concern classification of low, medium and high congestion labels within that dataset; they do not show that the model was used on live city traffic or on a road in Saudi Arabia. The researchers say the study is limited to one benchmark dataset and needs further validation with larger real urban datasets and real-time streams. The journal marks it as an accepted article shared early, with the final version subject to edits.