qingsongedu/time-series-transformers-review

Organizes transformer architectures for time series into a taxonomy covering forecasting, classification, anomaly detection, and imputation tasks, with links to 100+ peer-reviewed papers from top venues (ICLR, NeurIPS, IJCAI). Highlights key innovations like decomposition-based attention (Autoformer), patching strategies (PatchTST), and frequency-domain enhancements (FEDformer) that address challenges in long-range dependencies and non-stationary patterns. Includes curated implementations and benchmark datasets to support reproducible research across forecasting horizons and multivariate time series scenarios.

2,968 stars. No commits in the last 6 months.