Li, Ruojian (Zhejiang University), Guo, Wentong (Zhejiang University), Qi, Hongsheng (Zhejiang University)

Two-Dimensional Traffic Status Imputation: A Spatial and Temporal Joint Learning Using Multi-Scale Information

Scheduled for presentation during the Regular Session "Traffic prediction and estimation II" (WeBT6), Wednesday, September 25, 2024, 16:10−16:30, Salon 14

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

Missing values imputation in traffic time series is a significant research endeavor within intelligent transportation systems. This study introduces a multi-task training approach for missing data imputation in traffic time series and observed reconstruction. The paper presents the Masked Imputation Task (MIT) Block as a novel traffic time series imputation approach. Given the substantial and intricate seasonality inherent in traffic time series, this paper employs Fast Fourier Variation to transform one-dimensional multi-scale time series into two-dimensional data, thereby representing the sequence information. Additionally, the MLP structure is utilized to capture sequence features across both Intraperiod Variation and Interperiod Variation. This paper proposes the utilization of the Diagonally-Masked Self-Attention mechanism for achieving the observation reconstruction task. The diagonal mask substantially improves the reconstruction capability of the self-attention mechanism when applied to time series. Subsequently, the paper conducts experiments using an open-source traffic dataset, demonstrating that the proposed model in this study outperforms state-of-the-art models significantly in the domain of time series analysis. The code will be made public after the article is accepted.