<p>The practical applications of time series analysis are found across diverse domains, underscoring its pervasive utility. In this paper, we focus on the cross-channel dependence in multivariate time series, which is a very important feature in time series. Previous researches have predominantly concentrated on discerning temporal dependency, typically treating inter-channel relationship with simplistic notions of promiscuity or independence. However, such approaches fall short of accurately capturing intricate dependency arising from complex relationship between channels. To address this limitation, We proposed a Two-Stream Transformer Network (TSTNet) capable of adaptively extracting feature information across channels and time dimensions through explicit and implicit flows, respectively. By introducing the concept of inter-channel correlation scores, the proposed framework categorizes the channels into three distinct categories: explicitly correlated, implicitly correlated, and irrelevant channels. This categorization can decouple the intricate dependency between channels, facilitating the modeling of information for each channel and their interwoven relationship. Subsequently, we use two parallel streams to capture the feature information of explicitly correlated and implicitly correlated channels, and fuse them in an adaptive manner. Sufficient experiments demonstrate that our proposed TSTNet achieves consistent state-of-the-art performance on nine real-world datasets for two mainstream tasks including long-term forecasting and imputation.</p>