ADMFormer: New Transformer Model for Traffic Forecasting
Researchers have introduced ADMFormer, an Adaptive-Decomposition Transformer designed for traffic forecasting, which incorporates Time-Varying Masked Spatial Attention. This model tackles two significant issues: the coexistence of stable periodic patterns and event-driven variations in traffic data, as well as the dynamic and sparse spatial relationships among nodes. ADMFormer implements a time-node adaptive gating mechanism to separate traffic signals into prevailing regularities and time- and node-specific residual fluctuations. Additionally, it utilizes time-varying masked spatial attention to minimize unnecessary interactions and reduce noise from comprehensive all-pairs attention. The research paper can be found on arXiv under ID 2605.25543.
Key facts
- ADMFormer is a Transformer model for traffic forecasting.
- It uses an adaptive-decomposition mechanism to handle heterogeneous temporal patterns.
- Time-varying masked spatial attention addresses dynamic and sparse spatial dependencies.
- The model decouples traffic signals into dominant regularities and residual fluctuations.
- The paper is published on arXiv with ID 2605.25543.
- The approach aims to improve accuracy in intelligent transportation systems.
- Existing methods often treat temporal patterns within a unified representation.
- Dense all-pairs attention introduces redundant interactions and amplifies noise.
Entities
Institutions
- arXiv