Spatiotemporal-aware Trend-Seasonality Decomposition Network for Traffic Flow Forecasting
Authors: Lingxiao Cao, Bin Wang, Guiyuan Jiang, Yanwei Yu, Junyu Dong
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments conducted on real-world traffic datasets demonstrate that STDN achieves superior performance with remarkable computation cost. Furthermore, we have released a new traffic dataset named Ji Nan, which features unique inner-city dynamics, thereby enriching the scenario comprehensiveness in traffic prediction evaluation. |
| Researcher Affiliation | Academia | Faculty of Information Science and Engineering, Ocean University of China EMAIL, EMAIL |
| Pseudocode | No | The paper describes the methodology using mathematical equations and textual explanations of modules, but it does not include a clearly labeled pseudocode block or algorithm. |
| Open Source Code | Yes | All source code and data are available at https://github.com/roarer008/STDN |
| Open Datasets | Yes | To evaluate the performance of STDN, we utilize three real-world datasets, each offering unique traffic flow dynamics. The dataset Ji Nan, which is first released by us, derived from actual traffic flow statistics in a city, mirrors the setup in (Song et al. 2020), where the time interval is set to 5 minutes. Detailed descriptions of these datasets are provided in Table 1. |
| Dataset Splits | Yes | The datasets are split in a 6:2:2 ratio for training, validation, and testing, respectively. |
| Hardware Specification | Yes | Our experiments are conducted on a server with NVIDIA RTX 4090 GPU cards, running CUDA version 12.2. |
| Software Dependencies | Yes | Our experiments are conducted on a server with NVIDIA RTX 4090 GPU cards, running CUDA version 12.2. All the models are implemented using Py Torch. |
| Experiment Setup | Yes | We use Adam optimizer with an initial learning rate of 0.001. The standard batch size for all experiments is set to 64. If GPU memory constraints occur, the batch size is reduced to 32, and further to 16 if necessary, until the programs can run efficiently. The number of dimensions of node attribute on three datasets is C = 1. Totally, there are 3 hyperparameters in our model, i.e., the numbers of bottleneck transformer block L, the number of attention heads h, and the dimensionality d of each attention head, where the total number of features D = h d. The optimal settings for our model on Pe MS04 and Ji Nan datasets are L = 2, h = 8, d = 16 (D = 128). For the Pe MS07 dataset, the best performance is achieved with L = 2, h = 8, d = 12 (D = 96). |