Non-stationary Diffusion For Probabilistic Time Series Forecasting
Authors: Weiwei Ye, Zhuopeng Xu, Ning Gui
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments conducted on nine real-world and synthetic datasets demonstrate the superior performance of Ns Diff compared to existing approaches. |
| Researcher Affiliation | Academia | 1School of Computer Science and Engineering, Central South University, Changsha, China. |
| Pseudocode | Yes | Algorithm 1 Training Algorithm 2 Inference |
| Open Source Code | Yes | Code is available at https: //github.com/wwy155/Ns Diff. |
| Open Datasets | Yes | Nine popular real-world datasets with diverse characteristics are selected, including Electricity (ECL), ILI, ETT{h1, h2, m1, m2}, Exchange Rage (EXG), Traffic, and Solar Energy (Solar). 1https://archive.ics.uci.edu/ml/datasets/Electricity Load Diagrams20112014 2https://gis.cdc.gov/grasp/fluview/fluportaldashboard.html 3http://pems.dot.ca.gov/ 4http://www.nrel.gov/grid/solar-power-data.html |
| Dataset Splits | Yes | For dataset splits, we follow previous time series prediction works (Wu et al., 2022; Li et al., 2024b): the ETT datasets are split 12/4/4 months for train/val/test, while others are split 7:1:2. |
| Hardware Specification | No | The paper does not explicitly describe the specific hardware used to run its experiments. |
| Software Dependencies | No | The paper mentions using PyTorch and Numpy but does not provide specific version numbers for these or any other software dependencies. |
| Experiment Setup | Yes | All experiments are run with seeds {1, 2, 3} for 10 epochs. We use the best result from the validation set to evaluate the model on the test set. The learning rate is set to 0.001, batch size of 32 and the number of timesteps T = 20, consistent with prior work (Rasul et al., 2021). We employ a linear noise schedule with β1 = 10 4 and βT = 0.02, in line with the setup used in conventional DDPM (Ho et al., 2020). |