ElasTST: Towards Robust Varied-Horizon Forecasting with Elastic Time-Series Transformer

Authors: Jiawen Zhang, Shun Zheng, Xumeng Wen, Xiaofang Zhou, Jiang Bian, Jia Li

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Through comprehensive experiments and comparisons with state-of-the-art time-series architectures and contemporary foundation models, we demonstrate the efficacy of Elas TST s unique design elements.
Researcher Affiliation Collaboration Jiawen Zhang DSA, HKUST(GZ) Guangzhou, China jiawe.zh@gmail.com Shun Zheng Microsoft Research Asia Beijing, China shun.zheng@microsoft.com Xumeng Wen Microsoft Research Asia Beijing, China xumengwen@microsoft.com Xiaofang Zhou CSE, HKUST Hong Kong SAR, China zxf@ust.hk Jiang Bian Microsoft Research Asia Beijing, China jiang.bian@microsoft.com Jia Li DSA, HKUST(GZ) Guangzhou, China jialee@ust.hk
Pseudocode No The paper includes an architectural overview diagram (Figure 1) but does not provide any explicit pseudocode or algorithm blocks.
Open Source Code Yes Elas TST is open-sourced at https://github.com/microsoft/Prob TS/tree/elastst.
Open Datasets Yes Our experiments leverage 8 well-recognized datasets, including 4 from the ETT series (ETTh1, ETTh2, ETTm1, ETTm2), and others include Electricity, Exchange, Traffic, and Weather. These datasets cover a wide array of real-world scenarios and are commonly used as benchmarks in the field. Detailed descriptions of each dataset are provided in Appendix C.1. Datasets are available at https://github.com/thuml/Autoformer under MIT License.
Dataset Splits No The paper mentions 'NMAE metric used for model checkpointing' implying the use of a validation set, but it does not explicitly provide the training/validation/test dataset splits (e.g., percentages or sample counts) within the paper. It refers to common practices in long-term forecasting [33].
Hardware Specification Yes experiments are conducted on NVIDIA Tesla V100 GPUs with CUDA 12.1.
Software Dependencies No Elas TST is implemented using Py Torch Lightning [12]. The paper mentions this framework but does not provide specific version numbers for PyTorch Lightning or other key software libraries.
Experiment Setup Yes Elas TST is implemented using Py Torch Lightning [12], with a training regimen of 100 batches per epoch, a batch size of 32, and a total duration of 50 epochs. We use the Adam optimizer with a learning rate of 0.001... The range and specifics of these hyperparameters are documented in Appendix C.2.