Multi-Resolution Diffusion Models for Time Series Forecasting

Authors: Lifeng Shen, Weiyu Chen, James Kwok

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results on nine real-world time series datasets demonstrate that mr-Diff outperforms state-of-the-art time series diffusion models. It is also better than or comparable across a wide variety of advanced time series prediction models. In this section, we conduct time series prediction experiments by comparing 22 recent strong prediction models on 9 popular real-world time series datasets.
Researcher Affiliation Academia Lifeng Shen1 , Weiyu Chen2, James T. Kwok2 1 Division of Emerging Interdisciplinary Areas, Hong Kong University of Science and Technology 2 Department of Computer Science and Engineering, Hong Kong University of Science and Technology
Pseudocode Yes The pseudocode for the training and inference procedures of the backward denoising process can be found in Appendix A. Algorithm 1 Training procedure of the backward denoising process. Algorithm 2 Inference procedure.
Open Source Code No The paper does not provide an explicit statement or link for the open-source code of their proposed method.
Open Datasets Yes Experiments are performed on nine commonly-used real-world time series datasets (Zhou et al., 2021; Wu et al., 2021; Fan et al., 2022): (i) Nor Pool 1, (ii) Caiso 2, (iii) Traffic 3, (iv) Electricity 4, (v) Weather 5, (vi) Exchange 6 (Lai et al., 2018), (vii)-(viii) ETTh1 and ETTm1,7, (ix) Wind 8 (Li et al., 2022).
Dataset Splits No The paper mentions using a validation set to select history length but does not specify the exact percentages or counts for training, validation, or test dataset splits.
Hardware Specification Yes All experiments are run on an Nvidia RTX A6000 GPU with 48GB memory.
Software Dependencies No The paper mentions using Adam for training but does not provide specific version numbers for any software dependencies like programming languages or libraries.
Experiment Setup Yes We train the proposed model using Adam with a learning rate of 10 3. The batch size is 64, and training with early stopping for a maximum of 100 epochs. K = 100 diffusion steps are used, with a linear variance schedule (Rasul et al., 2021) starting from β1 = 10 4 to βK = 10 1. S = 5 stages are used. The history length (in {96, 192, 336, 720, 1440}) is selected by using the validation set.