A Large-Scale Spatio-Temporal Multimodal Fusion Framework for Traffic Prediction

Abstract

Traffic prediction is crucial for urban planning and transportation management, and deep learning techniques have emerged as effective tools for this task. While previous works have made advancements, they often overlook comprehensive analyses of spatio-temporal distributions and the integration of multimodal representations. Our research addresses these limitations by proposing a large-scale spatio-temporal multimodal fusion framework that enables accurate predictions based on location queries and seamlessly integrates various data sources. Specifically, we utilize Convolutional Neural Networks (CNNs) for spatial information processing and a combination of Recurrent Neural Networks (RNNs) for final spatio-temporal traffic prediction. This framework not only effectively reveals its ability to integrate various modal data in the spatio-temporal hyperspace, but has also been successfully implemented in a real-world large-scale map, showcasing its practical importance in tackling urban traffic challenges. The findings presented in this work contribute to the advancement of traffic prediction methods, offering valuable insights for further research and application in addressing real-world transportation challenges.