JCCMTM: Joint channel-independent and channel-dependent strategy for masked multivariate time-series modeling

Open Webpage

Qi Li, Zhenyu Zhang, Yong Zhang, Zhao Zhang, Lin Zhu, Xiaolei Hua, Renkai Yu, Xinwen Fan, Zhe Lei, Junlan Feng

Published: 01 Jan 2025, Last Modified: 28 Sept 2025Neural Networks 2025EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: Highlights•A novel MTM framework that explicitly models cross-series dependency in MTS data.•A novel paradigm to aggregate masked token embeddings from two MTS modeling strategies.•Develop a dual-stream self-attention-based backbone to jointly model intra- and cross-series dependencies.•Optimization schemes to significantly reduce the model’s computational complexity.