Shengmin Guo

Shengmin Guo

Auto-generated profile, based on publicly available information

Shengmin Guo

Beijing PalmGo Infotech Company Ltd., Beijing, China

Research Focus & Metrics

Academic Career Length

1Year (2026-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
6 February 2026 in IEEE Transactions on Intelligent Transportation Systems
Verified
Journal Article
Closed access
XLXiaoyue Luo
Xiaoyue Luo
CSCheng Shifen
Cheng Shifen
YYYitao Yang
Yitao Yang
PWPeixiao Wang
Peixiao Wang
SGShengmin Guo
Shengmin Guo
Feng LuFeng Lu
Feng Lu
Tensor decomposition has been widely recognized as a promising approach for addressing missing data in traffic flow analysis. Recent studies have demonstrated its effectiveness in capturing complex spatiotemporal patterns and improving imputation accuracy. However, two challenges still remain: i)The spatiotemporal heterogeneity at missing locations is difficult to identify and accurately represent, preventing the model from optimizing based on the true data distribution; ii) The lack of a dynamic feedback mechanism hinders the model to adaptively optimize its spatiotemporal representation based on imputation results, thus limiting the generalization and robustness. This paper proposes a heterogeneity-guided tensor decomposition method (Hg-TD) to address these problems. First, we design a spatiotemporal-decoupled masking and encoding strategy, leveraging contrastive learning based on reconstruction loss to generate a dynamic spatial weight matrix, thereby enhancing the representation of spatiotemporal heterogeneity at missing locations. Second, we introduce a dynamic feedback-driven self-supervised learning mechanism that uses imputation results as feedback to dynamically optimize both representation learning and tensor decomposition, allowing the model to adaptively adjust spatiotemporal constraints, thereby improving imputation performance under sparse and non-stationary data conditions. Experimental results on four traffic flow datasets from urban areas demonstrate that Hg-TD consistently surpasses fifteen baseline models across various data missing patterns, and achieves a favorable balance between computational complexity and imputation performance. The improvement relative to the baselines is particularly pronounced under conditions of severe data missingness. Codes and datasets are available at https://figshare.com/s/17b6beeb8e8bad5e3ee7