Spatial-Temporal Feedback Diffusion Guidance for Controlled Traffic Imputation

Xiaowei Mao, Huihu Ding, Yan Lin, Tingrui Wu, Shengnan Guo, Dazhuo Qiu, Feiling Fang, Jilin Hu, Huaiyu Wan · 2026

Read Spatial-Temporal Feedback Diffusion Guidance for Controlled Traffic Imputation by Xiaowei Mao, Huihu Ding, Yan Lin, Tingrui Wu, Shengnan Guo, Dazhuo Qiu, Feiling Fang, Jilin Hu, Huaiyu Wan (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-04572.txt.

Description

Xiaowei Mao and eight co-authors fill gaps in traffic data with a diffusion model that listens to feedback, in a 2026 arXiv paper. Sensors fail constantly, and the missing readings hobble transport systems that plan around them. Diffusion-based imputers generate the missing values by reversing a noising process, steered by the values that were observed. they steer with one guidance strength everywhere, which starves the worst-hit locations, where sparse readings cannot hold the generation on course.

Their method, FENCE, adjusts the steering dynamically: guidance strengthens when generated values drift from observations and relaxes as alignment improves, preventing overcorrection. Because alignment differs by place, guidance is also set per cluster of locations, grouped by their attention patterns so spatial and temporal correlations inform the scale. On real traffic datasets the feedback scheme significantly improved imputation accuracy over existing diffusion approaches.

Catalog record

CREATORXiaowei Mao, Huihu Ding, Yan Lin, Tingrui Wu, Shengnan Guo, Dazhuo Qiu, Feiling Fang, Jilin Hu, Huaiyu Wan
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE59 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-04572

Browse the full library · About the Tower of Babel