<ul data-eligibleForWebStory="true"><li>Self-supervised learning in time series analysis is gaining attention for reducing the need for labeled data and improving downstream tasks.</li><li>Current methods struggle to capture both long-term dynamic evolution and subtle local patterns effectively.</li><li>A new model called TimeDART is introduced, which unifies two generative paradigms for learning transferable representations.</li><li>TimeDART uses a causal Transformer encoder and patch-based embedding strategy to capture evolving trends from left to right.</li><li>The model also employs a denoising diffusion process to capture fine-grained local patterns through forward diffusion and reverse denoising.</li><li>Optimization of the model is done in an autoregressive manner, effectively combining global and local sequence features.</li><li>Extensive experiments on public datasets show that TimeDART outperforms existing methods in time series forecasting and classification tasks.</li><li>The code for TimeDART is available at https://github.com/Melmaphother/TimeDART.</li></ul>

TimeDART: A Diffusion Autoregressive Transformer for Self-Supervised Time Series Representation

Discover more