Learning Latent Seasonal-Trend Representations for Time Series Forecasting

Authors: Zhiyuan Wang, Xovee Xu, Weifeng Zhang, Goce Trajcevski, Ting Zhong, Fan Zhou

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments prove that La ST achieves state-of-the-art performance on time series forecasting task against the most advanced representation learning and end-to-end forecasting models.
Researcher Affiliation Academia 1 University of Electronic Science and Technology of China 2 Iowa State University
Pseudocode Yes Algorithm 1 An epoch of the optimization of La ST.
Open Source Code Yes For reproducibility, our implementation is publicly available on Github1. ... 1https://github.com/zhycs/La ST
Open Datasets Yes We conducted our experiments on seven real-world benchmark datasets from four categories of mainstream time series forecasting applications: (1) ETT 2[25]: Electricity Transformer Temperature... (2) Electricity, from the UCI Machine Learning Repository 3 and preprocessed by [56]... (3) Exchange [56]... (4) Weather 4... Footnotes provide URLs: 2https://github.com/zhouhaoyi/ETDataset, 3https://archive.ics.uci.edu/ml/datasets/Electricity Load Diagrams, 4https://www.bgc-jena.mpg.de/wetter
Dataset Splits Yes For the dataset split, we follow a standard protocol that categorizes all datasets into training, validation, and test set in chronological order by the ratio of 6:2:2 for all datasets.
Hardware Specification Yes All experiments were conducted on a server with an NVIDIA RTX 3090 GPU and an Intel(R) Core(TM) i9-10900K CPU @ 3.70GHz. The system had 128 GB of RAM.
Software Dependencies No The paper mentions using the Adam optimizer, but does not provide specific version numbers for any software dependencies like programming languages, libraries, or frameworks (e.g., Python, PyTorch, TensorFlow versions).
Experiment Setup Yes We set them as 32 in univariate forecasting and as 128 in multivariate forecasting on other datasets. MAE loss is used to measure the forecasting derived from the predictor. For the training strategy, we use the Adam [57] optimizer, and training process is early stopped within 10 epochs. We initialize the learning rate with 10-3 and decay it with 0.95 weight every epoch.