Frequency-domain MLPs are More Effective Learners in Time Series Forecasting

Authors: Kun Yi, Qi Zhang, Wei Fan, Shoujin Wang, Pengyang Wang, Hui He, Ning An, Defu Lian, Longbing Cao, Zhendong Niu

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on 13 real-world benchmarks (including 7 benchmarks for short-term forecasting and 6 benchmarks for long-term forecasting) demonstrate our consistent superiority over state-of-the-art methods.
Researcher Affiliation Academia 1Beijing Institute of Technology, 2Tongji University, 3University of Oxford 4University of Technology Sydney, 5University of Macau, 6USTC 7He Fei University of Technology, 8Macquarie University
Pseudocode No The paper does not contain any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes Code is available at this repository: https://github.com/aikunyi/Fre TS.
Open Datasets Yes Our empirical results are performed on various domains of datasets, including traffic, energy, web, traffic, electrocardiogram, and healthcare, etc. Specifically, for the task of short-term forecasting, we adopt Solar 2, Wiki [37], Traffic [37], Electricity 3, ECG [16], METR-LA [38], and COVID-19 [4] datasets, following previous forecasting literature [16].
Dataset Splits Yes We split the datasets into training, validation, and test sets by the ratio of 7:2:1 except for the COVID-19 datasets with 6:2:2.
Hardware Specification Yes Our model is implemented with Pytorch 1.8 [41], and all experiments are conducted on a single NVIDIA RTX 3080 10GB GPU.
Software Dependencies Yes Our model is implemented with Pytorch 1.8 [41], and all experiments are conducted on a single NVIDIA RTX 3080 10GB GPU.
Experiment Setup Yes By default, both the frequency channel and temporal learners contain one layer of Fre MLP with the embedding size d of 128, and the hidden size dh is set to 256. For short-term forecasting, the batch size is set to 32 for Solar, METR-LA, ECG, COVID-19, and Electricity datasets. And for Wiki and Traffic datasets, the batch size is set to 4. For the long-term forecasting, except for the lookback window size, we follow most of the experimental settings of LTSF-Linear [35]. The lookback window size is set to 96 which is recommended by FEDformer [30] and Autoformer [14]. We take MSE (Mean Squared Error) as the loss function and report MAE (Mean Absolute Errors) and RMSE (Root Mean Squared Errors) results as the evaluation metrics.