Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
DecompNet: Enhancing Time Series Forecasting Models with Implicit Decomposition
Authors: Donghao Luo, Xue Wang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we conduct extensive experiments to evaluate the effectiveness of our Decomp Net. In particular, we investigate the enhancement capability of Decomp Net by integrating it with various advanced time series backbone models in Section 4.1, compare our Decomp Net with other enhancement frameworks in Section 4.2 and compare our idea of implicit decomposition with the previous idea of explicit decomposition in Section 4.3. |
| Researcher Affiliation | Academia | Donghao Luo, Xue Wang Department of Precision Instrument, Tsinghua University, Beijing 100084, China EMAIL, EMAIL |
| Pseudocode | No | The paper describes its methodology in text and illustrates concepts with figures like "Figure 1: (a) Comparison between implicit decomposition and explicit decomposition. (b) Method overview of our Decomp Net." and "Figure 3: Illustration of Seasonal-Trend Re-parameterization.". However, there are no explicit "Pseudocode" or "Algorithm" blocks or figures. |
| Open Source Code | Yes | Code is available at this repository: https://github.com/luodhhh/Decomp Net. |
| Open Datasets | Yes | We conduct long-term forecasting experiments on 8 popular real-world benchmarks, including Weather [40], Traffic [33], ECL [36], Solar-Energy [29] and 4 ETT datasets [48]. More details of these datasets are provided in Appendix B. These datasets have been extensively utilized for benchmarking and cover many aspects of life. More introduction of the datasets are as follows: 1) Weather1 contains 21 meteorological indicators of Germany in 2020. 2) Traffic2 contains the road occupancy rates measured by 862 different sensors on San Francisco Bay area freeways in 2 years. 3) ECL(Electricity)3 contains hourly electricity consumption of 321 clients from 2012 to 2014. 4) ETT(Electricity Transformer Temperature)4 contains the data collected from two different electricity transformers (labeled as 1 and 2) with two different resolutions (denoted as m for 15 minutes and h for 1 hour) by 7 sensors. As a result, in total we have 4 ETT datasets: ETTh1, ETTh2, ETTm1, ETTm2. 5) Solar(Solar-Energy)5 contains 137 time series representing the solar power production in Alabama state in 2006. The links for these datasets are also provided in the footnotes. |
| Dataset Splits | Yes | We follow standard protocol [48] and split all datasets into training, validation and test set in chronological order by the ratio of 6:2:2 for the ETT datasets and 7:1:2 for the other datasets. And training, validation and test sets are zero-mean normalized with the mean and standard deviation of training set. |
| Hardware Specification | Yes | All the deep learning networks are implemented in Py Torch[32] and conducted on NVIDIA A100 40GB GPU. |
| Software Dependencies | No | All the deep learning networks are implemented in Py Torch[32] and conducted on NVIDIA A100 40GB GPU. While PyTorch is mentioned, a specific version number is not provided in the text or appendix for the implementation itself. |
| Experiment Setup | Yes | Following the previous settings, we set prediction lengths as {96, 192, 336, 720} and fix the input length as 384. We calculate the mean squared error (MSE) and mean absolute error (MAE) of multivariate time series forecasting as metrics. In all experiments, we set the moving average window size as 25 for time series decomposition during the decoupled pretraining stage. Our method is trained with the L2 loss, using the ADAM optimizer with an initial learning rate in {10 3, 5 10 4, 10 4}. The default training process is 30 epochs with proper early stopping (i.e., we set the maximum epoch as 30 and set the early stop patience as 3). More datasets details and implementation details are in Appendix B and C. |