TY - GEN
T1 - STMAE
T2 - 27th International Conference on Pattern Recognition, ICPR 2024
AU - Wu, Xing
AU - Cai, Chengyou
AU - Wang, Xiaoxiao
AU - Wang, Jianjia
AU - Yao, Junfeng
AU - Qian, Quan
AU - Song, Jun
N1 - Publisher Copyright:
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2025.
PY - 2025
Y1 - 2025
N2 - The advancement of intelligent transportation systems underscores the importance of data-driven approaches in traffic forecasting, which plays a crucial role in tasks such as traffic signal control and route guidance, among others. However, the inherent uncertainty stemming from regional traffic dynamics, coupled with intricate spatio-temporal correlations, poses formidable challenges to accurate traffic prediction. Moreover, the complexities inherent in sequence forecasting across varying scales further exacerbate the accuracy dilemma. Recognizing the need for integrating information across spatial and temporal dimensions to enhance forecasting precision, a novel solution termed Spatial Temporal Masked Autoencoder (STMAE) is introduced. The STMAE framework addresses these challenges through a two-stage learning process. In the pre-training phase, an autoencoder architecture is employed to extract spatio-temporal features from the data. In the fine-tuning phase, the pre-trained encoder of the STMAE model undergoes further refinement to specifically target traffic forecasting tasks. Extensive evaluations validate the effectiveness of the proposed STMAE model. Notably, STMAE demonstrates competitive performance, achieving 3.32 Vehs MAE for long-term (60 min) traffic forecasting while operating within a reduced computational budget.
AB - The advancement of intelligent transportation systems underscores the importance of data-driven approaches in traffic forecasting, which plays a crucial role in tasks such as traffic signal control and route guidance, among others. However, the inherent uncertainty stemming from regional traffic dynamics, coupled with intricate spatio-temporal correlations, poses formidable challenges to accurate traffic prediction. Moreover, the complexities inherent in sequence forecasting across varying scales further exacerbate the accuracy dilemma. Recognizing the need for integrating information across spatial and temporal dimensions to enhance forecasting precision, a novel solution termed Spatial Temporal Masked Autoencoder (STMAE) is introduced. The STMAE framework addresses these challenges through a two-stage learning process. In the pre-training phase, an autoencoder architecture is employed to extract spatio-temporal features from the data. In the fine-tuning phase, the pre-trained encoder of the STMAE model undergoes further refinement to specifically target traffic forecasting tasks. Extensive evaluations validate the effectiveness of the proposed STMAE model. Notably, STMAE demonstrates competitive performance, achieving 3.32 Vehs MAE for long-term (60 min) traffic forecasting while operating within a reduced computational budget.
KW - Data-driven Methods
KW - Masked Autoencoder
KW - Traffic Forecasting
UR - http://www.scopus.com/inward/record.url?scp=85211367893&partnerID=8YFLogxK
U2 - 10.1007/978-3-031-78169-8_14
DO - 10.1007/978-3-031-78169-8_14
M3 - Conference Proceeding
AN - SCOPUS:85211367893
SN - 9783031781681
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 209
EP - 223
BT - Pattern Recognition - 27th International Conference, ICPR 2024, Proceedings
A2 - Antonacopoulos, Apostolos
A2 - Chaudhuri, Subhasis
A2 - Chellappa, Rama
A2 - Liu, Cheng-Lin
A2 - Bhattacharya, Saumik
A2 - Pal, Umapada
PB - Springer Science and Business Media Deutschland GmbH
Y2 - 1 December 2024 through 5 December 2024
ER -