Huang, Shouyu (Beijing University of Posts and Telecommunications), Wang, Luhan (Beijing University of Posts and Telecommunications), Zheng, Rui (Beijing Jiaotong University), Huo, Jie (Beijing University of Posts and Telecommunications), Wen, Xiangming (Beijing University of Posts and Telecommunications), Lu, Zhaoming (Beijing University of Posts and Telecommunications)

Multi-Attention Adaptive Graph Transformer Network for Data Inference and Prediction in Vehicular Crowdsensing

Scheduled for presentation during the Poster Session "Data Mining and Data Analysis" (FrAT14), Friday, September 27, 2024, 10:30−12:30, Foyer

2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC), September 24- 27, 2024, Edmonton, Canada

This information is tentative and subject to change. Compiled on December 26, 2024

Abstract

Vehicular crowdsensing has emerged as an important sensing paradigm in the Internet of Things (IoT), which recruits Intelligent Vehicles to collect data. As its variant, sparse MCS collects data from some sensing areas and infer data for other unsensed areas, addressing the challenge of maintaining high-quality sensing at manageable costs. In practical vehicular crowdsensing scenarios, the diverse requirements of urban applications make it crucial not only to infer data for the current period but also to predict the entire sensing map for the future. This predictive capability not only reduces the cost of sensing but also provides critical data support for various urban applications, including intelligent transportation systems. This paper introduces the Multi-Attention Adaptive Graph Transformer Network (MAAGTN) aimed at improving the accuracy of data inference and prediction in vehicular crowdsensing. We present a novel embedding method that integrates spatial-temporal information of sensing maps into the model. An adaptive learnable graph module and a multi-attention module are designed to dynamically capture complex spatial-temporal correlations among data. Furthermore, a dynamic multi-task learning framework is proposed to mitigate error propagation by adjusting task weights during training. Experimental evaluations on real-world datasets demonstrate the superiority of MAAGTN for data inference and prediction.