SimMTM
Summary
SimMTM is a masked time-series modeling framework that reconstructs an original series from multiple masked neighbors and learns neighborhood structure in representation space.
Role In The Wiki
SimMTM is an important masked-modeling baseline because it attacks a time-series-specific failure mode: random masking can destroy the temporal variations that carry the signal. Its multi-neighbor reconstruction design is a useful contrast to ordinary masked patch reconstruction and latent-prediction methods.
Official Artifacts
- Official code: https://github.com/thuml/SimMTM
- Checkpoint archive: https://cloud.tsinghua.edu.cn/f/466995bb5f924f55a6da/?dl=1