Modality Matters: Universal Time Series Modeling via Channel Dependency Search

Minjun Cao, Hao Miao, Wentao Zhang, Senzhang Wang

Keywords: Time Series Modeling, Channel Dependency Search

Abstract: The expanding development of wireless and mobile devices results in a proliferation of multivariate time series data, enabling various analytical tasks, e.g., forecasting, classification, and anomaly detection. Most existing time series modeling methods are dedicated to developing task-specific models due to the heterogeneous dimensionalities, resulting in inefficient resource utilization and limited cross-domain transferability. To address this issue, this study achieves a unified paradigm transcending task boundaries and proposes a universal modality-aware Time series modeling framework leveraging Channel Dependency Search named TimeCDS. Specifically, TimeCDS innovatively identifies a certain number of representative features by projecting the heterogeneous time series features into the hierarchical spaces and dynamically modeling their inter-channel relationships to alleviate the heterogeneity issue. A novel time series imaging method is then proposed to automatically introduce the image modality from sequences, facilitating the comprehensive temporal-spatial pattern extraction. Further, a dual-branch architecture is designed to process the sequential data and the visual representations simultaneously, exploiting the complementary cross-modal features through the proposed Cross-Modal Attention and Dynamic Weighted-Averaging. Extensive experiments across different analytical tasks demonstrate the consistently superior performance of TimeCDS, outperforming existing state-of-the-art baselines by up to 15.9%. The code of TimeCDS is publicly available at https://anonymous.4open.science/r/TimeCDS/.

Primary Area: learning on time series and dynamical systems

Submission Number: 16534