Application Research of Cross-Attention Mechanism for Traffic Prediction Based on Heterogeneous Data

Faculty of Data Science, City University of Macau, Macau, 999078, China

Corresponding author: D22090101344@cityu.edu.mo

Abstract

Intelligent transportation systems need to be developed with precise traffic flow predictions to reduce traffic accidents, improve overall urban mobility, and mitigate congestion. The intricacy and variety of traffic conditions are often too complex and variable for traditional approaches to handling, especially when dealing with heterogeneous event data from several sources like weather variations and traffic incidents. This review highlights the significance of cross-attention mechanisms by examining the developments in integrating multi-source heterogeneous event data for traffic prediction. Examining different approaches used in previous work, the study focuses on the Event-aware Graph Attention Fusion Network (EGAF-Net). This cutting-edge model efficiently integrates and analyzes complex spatial-temporal data. Through an analysis of these methods, the research demonstrates how applying advanced deep learning algorithms and cross-attention processes has significantly improved prediction robustness and accuracy. The results underscore the critical role of heterogeneous data integration in enhancing traffic prediction models, providing insights into current challenges and potential future developments in the field. This thorough analysis aims to direct future research endeavors and open the door for more dependable and effective intelligent transportation systems.