Reproduce “DiffStyleTS: Diffusion Model for Style Transfer in Time Series” on OpenResearch

About this paper

Mayank Nagda, Phil Ostheimer, Justus Arweiler, Indra Jungjohann, Jennifer Werner, Dennis Wagner, Aparna Muraleedharan, Pouya Jafari and 8 more

DiffStyleTS, a novel diffusion-based framework, introduces the first application of diffusion models to time series style transfer. It effectively disentangles the content and style of time series to generate diverse and realistic new sequences, outperforming existing baselines and demonstrating utility in low-data anomaly detection scenarios.

Key insights

• Diffusion models can be effectively adapted for time series style transfer by incorporating inductive biases in specialized encoders that disentangle content (global structure) and style (local dynamics) across different frequency domains.
• The use of a conditional Diffusion Transformer, with separate cross-attention for content and style, allows for fine-grained control over the generation process and the synthesis of diverse outputs.
• Attention with Linear Biases (ALiBi) enables robust zero-shot generalization of style transfer to time series lengths significantly greater than those encountered during training.

Summary by alphaXiv.