When and Where to Go Next: Deep Learning Framework for Modeling Drivers’ Behaviors Using Automatic Vehicle Identification Data

• Published in: Transportation research record Vol. 2676; no. 6; pp. 387 - 398
• Main Authors: Jin, Kun, Li, Xinran, Wang, Wei, Hua, XueDong, Qin, Shaoyang
• Format: Journal Article
• Language: English
• Published: Los Angeles, CA SAGE Publications 01.06.2022
• Subjects: spatial data; deep learning; data and data science; trajectory
• ISSN: 0361-1981; 2169-4052
• DOI: 10.1177/03611981221074372

In recent years, automatic vehicle identification (AVI) systems have rapidly developed in many countries, which provides an excellent opportunity to understand drivers’ mobility patterns in urban road networks. Few efforts have been devoted to fully utilizing the AVI data to address the prediction issue in behaviors modeling. This paper proposes a deep learning (DL) framework driven by AVI data to model drivers’ behaviors and further incorporate travel time prediction in the next location prediction problem. Specifically, DeepWalk encoder and DeepWalk + Time long short-term memory (DT-LSTM) were proposed to capture the spatial and temporal correlations simultaneously. By learning the spatial relationship between the sensors from the historical trajectories, the DeepWalk encoder converted the sensors into low-dimensional numerical vectors essential for DL. As a vector form, the spatial relationship of the sensor can be adequately measured. Besides, a new long short-term memory (LSTM) variant, DT-LSTM, was designed to memorize drivers’ short-term and long-term interests, respectively. By adding additional time gates, DT-LSTM facilities the location and travel time consistency in the trajectories sequences with different time intervals. Experimental results demonstrated the effectiveness of the model, which achieved the state-of-the-art performance (SOTA) compared with other benchmark methods, with the Acc-1 of 82%, Acc-5 of 95%, and MAPE of 17%.