Arrow: Accelerator for Time Series Causal Discovery with Time Weaving
Authors: Yuanyuan Yao, Yuan Dong, Lu Chen, Kun Kuang, Ziquan Fang, Cheng Long, Yunjun Gao, Tianyi Li
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We applied ARROW to four different types of time series causal discovery algorithms and evaluated it on 25 synthetic and real-world datasets. The results demonstrate that, compared to the original algorithms, ARROW achieves up to 153x speedup while achieving higher accuracy in most cases. |
| Researcher Affiliation | Academia | 1The College of Computer Science, Zhejiang University, Hangzhou 310027, China 2College of Computing and Data Science, Nanyang Technological University, Singapore 3The Department of Computer Science, Aalborg University, Denmark. Correspondence to: Lu Chen <EMAIL>. |
| Pseudocode | Yes | Algorithm 1 Candidates Pruning |
| Open Source Code | Yes | The source code of ARROW is available 2. https://github.com/Xiangguan Mu/arrow |
| Open Datasets | Yes | In addition to the synthetic datasets, we also conduct validation on the real-world Dream3 1 dataset. 1https://www.synapse.org/Synapse:syn3033083/files/ |
| Dataset Splits | No | The paper mentions generating synthetic datasets with specific characteristics (e.g., "10 variables and a time length of 1000"), and using predefined ranges for time lags. However, it does not provide explicit training/validation/test splits, percentages, or absolute counts for any of the datasets used. |
| Hardware Specification | Yes | All methods are executed on a machine equipped with an Intel(R) Core(TM) i9-10900K CPU, boasting 10 cores and a clock speed of 3.70GHz. The system also features an NVIDIA Ge Force RTX 3090 graphics card, equipped with 24GB of video memory. |
| Software Dependencies | No | The paper references open-source code for baselines (e.g., PCMCI, SURD, NGC, VARLiNGAM) by providing URLs to their repositories, but it does not specify version numbers for these software components or any other underlying libraries/frameworks (like Python, PyTorch, TensorFlow, etc.) used for implementation. |
| Experiment Setup | Yes | We synthesized a dataset with 10 variables and a time length of 1000. The window size w for time weaving was set to 1. In addition, we conducted experiments with constant time lags and multiple time lags. For the constant time lags, the lags between variables is fixed, and it can be selected from the set {3, 5, 7, 9, 15, 20}. In contrast, the multiple time lags represent varying lags between variables, with the lag value being chosen from the set {3, 5, 7, 9, 15, 20} as the range for the time lags. The experiments on varying time-lagged edges are deferred to Appendix D, while in the main experiments, we set k to {5, 15}. |