WaveletDiff framework generates synthetic time series using wavelet coefficients

Researchers have developed WaveletDiff, a novel framework for generating synthetic time series data by training diffusion models on wavelet coefficients. This approach leverages the multi-resolution structure of time series, incorporating transformers and cross-level attention mechanisms for information exchange across different scales. WaveletDiff also utilizes level-specific energy constraints derived from Parseval's theorem to maintain time-frequency properties during generation. Experiments on six real-world datasets show WaveletDiff outperforms existing diffusion models like FourierDiffusion and Diffusion-TS, and is comparable to MSDformer while using fewer parameters and less training time. AI

IMPACT This new method for synthetic time series generation could improve data availability for forecasting and analysis in fields like healthcare and finance.

RANK_REASON The cluster describes a new research paper detailing a novel framework for time series generation. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• Diffusion-TS
• FourierDiffusion
• MSDformer
• Parseval's theorem
• SigDiffusions
• WaveletDiff
• Yu-Hsiang Wang

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →