Tri-Tense Former: Capturing Dynamic Traffic Flow Using Tri-Tense Attention for Traffic Forecasting

Yoojin Kim, Kyeongmin Do, Jeong hwan Jeon

Keywords: Traffic forecasting, Spatio-temporal dependency, Transformer, Contrastive learning

TL;DR: TTformer captures spatio-temporal relationships through tense-specific attention modules.

Abstract:

Accurate traffic forecasting is essential to enable advanced utilization of intelligent transportation systems. However, forecasting models often struggle to capture the complex spatio-temporal dependencies of traffic data, as they typically handle spatial and temporal dependencies separately. To overcome this limitation, we introduce the Tri-Tense Former (TTformer), a novel approach that captures spatio-temporal relationships through three tense-specific attention modules. We categorize traffic flow into three tense dimensions: past-to-present (present-perfect), present, and future. Each tense-specific attention module captures the dependencies within its respective traffic flow. Furthermore, to address incomplete traffic data, we improve the robustness of the model by employing contrastive learning with negative filtering technique that operates regardless of predefined adjacency matrices. TTformer significantly outperforms existing models by more effectively capturing spatio-temporal dependencies and improving traffic forecasting accuracy.

Primary Area: learning on time series and dynamical systems

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 10134