AIMEN: A Multiexpert Network for Traffic Prediction With Accident Impact Learning

The advent of advanced sensor technologies has enabled the collection of rich traffic data in smart cities, paving the way for data-driven traffic forecasting services. These services play a vital role in optimizing traffic management, alleviating congestion, and enhancing transportation system efficiency. Although existing spatiotemporal graph neural networks (STGNNs) have made strides in traffic prediction by leveraging multiple graph perspectives, they often overlook the dynamic nature of traffic networks and fail to fully integrate the impact of traffic accidents. To address these limitations, we introduce the accident impact multiexpert network (AIMEN), which incorporates the accident impact learning expert (AILE) to encode detailed accident profiles and generate Accident Propagation Graphs that model how accidents influence traffic patterns across road networks. AIMEN also employs spatiotemporal learning experts (STLEs) to capture varying spatial relationships through distinct graph views, enhancing the model’s ability to adapt to changing traffic conditions. Experimental results on real-world datasets show that AIMEN outperforms state-of-the-art methods, reducing the root mean square error (RMSE) by 5.06% and the mean absolute percentage error (MAPE) by 4.76% compared to suboptimal results.