A Non-isotropic Time Series Diffusion Model with Moving Average Transitions
Authors: Chenxi Wang, Linxiao Yang, Zhixian Wang, Liang Sun, Yi Wang
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experiments on various datasets demonstrated MA-TSD s superior performance in time series forecasting and super-resolution tasks. ... We conducted extensive experiments to demonstrate our salient performances over existing DDPM-based diffusion models on time series-related tasks like time series forecasting and time series super-resolution. ... In this section, we mainly focus on two important time series analysis tasks, forecasting and super-resolution. The standard time series synthesis task is included in Appendix C.1. An ablation study of our framework design is also included. |
| Researcher Affiliation | Collaboration | 1Department of Electrical and Electronic Engineering, The University of Hong Kong, Hong Kong SAR 2DAMO Academy, Alibaba Group, Hangzhou, China. Correspondence to: Yi Wang <EMAIL>. |
| Pseudocode | No | The paper describes methods and processes through mathematical equations and textual descriptions, for example, the forward and backward processes in Section 4.2 and 4.3. While these describe algorithmic steps, there is no clearly labeled section or figure explicitly titled 'Pseudocode' or 'Algorithm'. |
| Open Source Code | Yes | Our codes can be found in https://github.com/Will Wang1113/Moving-Average-Diffusion. |
| Open Datasets | Yes | Datasets. We consider six real-world datasets with diverse temporal dynamics, commonly used by the community (Wang et al., 2024), namely Electricity, ETTh2, ETTm2, exchange, traffic, weather. ... We consider three high-resolution real-world datasets with 5-minute resolution, MFRED, Wind, Solar. ... For MFRED and Solar, the original resolution is 10 seconds and 1 minute respectively, we resampled them to 5 minutes for alignment. ... MFRED (Meinrenken et al., 2020), Wind7 and Solar8 are listed in Table 9. |
| Dataset Splits | No | The paper mentions training epochs with early stopping, which implies a validation set, but does not provide specific percentages or counts for training, validation, and test splits for any of the datasets used. For example, in Appendix C.2, it states: "The length of the look-back window is set as 96, and the target time series length is set as L {96, 192, 336, 720}. Such a setting is aligned with (Wang et al., 2024)." This defines input/output lengths but not the data splits themselves. |
| Hardware Specification | Yes | We launch our experiments on a single NVIDIA Ge Force RTX 4090 24GB GPU. |
| Software Dependencies | No | The paper does not provide specific version numbers for any software libraries or frameworks. It mentions: "The specific network architectures can be referred to as our source code.", implying details are in the code, but not explicitly stated in the paper. |
| Experiment Setup | Yes | We set the batch size as 64, the learning rate as 2 10 4, the training epoch as 100 with early stopping, and the diffusion step T = 100. ... The hyperparameters of hybrid optimization are chosen as λz = λµ = λσ = 1. |