Deciphering Spatio-Temporal Graph Forecasting: A Causal Lens and Treatment
Authors: Yutong Xia, Yuxuan Liang, Haomin Wen, Xu Liu, Kun Wang, Zhengyang Zhou, Roger Zimmermann
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments results on three real-world datasets demonstrate the effectiveness of Ca ST, which consistently outperforms existing methods with good interpretability. Our source code is available at https://github.com/yutong-xia/Ca ST. |
| Researcher Affiliation | Academia | Yutong Xia1, Yuxuan Liang2 , Haomin Wen3, Xu Liu1, Kun Wang4, Zhengyang Zhou4, Roger Zimmermann1 1National University of Singapore 2The Hong Kong University of Science and Technology (Guangzhou) 3Beijing Jiaotong University 4University of Science and Technology of China |
| Pseudocode | No | The paper describes algorithms and methods but does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our source code is available at https://github.com/yutong-xia/Ca ST. |
| Open Datasets | Yes | Datasets & Baselines. We conduct experiments using three real-world datasets (PEMS08 [42], AIRBJ [59], and AIR-GZ [59]) from two distinct domains to evaluate our proposed method. PEMS08 contains the traffic flow data in San Bernardino from Jul. to Aug. in 2016, with 170 detectors on 8 roads with a time interval of 5 minutes. AIR-BJ and AIR-GZ contain one-year PM2.5 readings collected from air quality monitoring stations in Beijing and Guangzhou, respectively. Our task is to predict the next 24 steps based on the past 24 steps. For comparison, we select two classical methods (HA [63] and VAR [45]) and seven state-of-the-art STGNNs for STG forecasting, including DCRNN [26], STGCN [61], ASTGCN[14], MTGNN[55], AGCRN[1], GMSDR [30] and STGNCDE [5]. More details about the datasets and baselines can be found in Appendix D and E, respectively. |
| Dataset Splits | Yes | Table 3: Statistics of datasets. Dataset #Nodes #Edges Data Type Time interval Date Range #Samples Train:Val:Test AIR-BJ 34 82 PM2.5 1 hour Jan.1, 2019 Dec. 31, 2019 8,760 4:1:1 AIR-GZ 41 77 PM2.5 1 hour Jan.1, 2017 Dec. 31, 2017 8,760 4:1:1 PEMS08 170 303 Traffic flow 5 minutes Jul. 1, 2016 Aug.31, 2016 17,856 8:1:1 |
| Hardware Specification | Yes | We implement Ca ST and baselines with Py Torch 1.10.2 on a server with NVIDIA RTX A6000. |
| Software Dependencies | Yes | We implement Ca ST and baselines with Py Torch 1.10.2 on a server with NVIDIA RTX A6000. |
| Experiment Setup | Yes | Implementation Details. We implement Ca ST and baselines with Py Torch 1.10.2 on a server with NVIDIA RTX A6000. We use the TCN [2] as the backbone encoder and a 3-layer MLP as the predictor and the classifier. Our model is trained using Adam optimizer [22] with a learning rate of 0.001 and a batch size of 64. For the hidden dimension F, we conduct a grid search over {8, 16, 32, 64}. For the number of layers in each convolutional block, we test it from 1 to 3. The codebook size K is searched over {5, 10, 20}. See the final setting of our model on each dataset in Appendix D. |