DisenTS: Disentangled Channel Evolving Pattern Modeling for Multivariate Time Series Forecasting

• DOI: 10.48550/arxiv.2410.22981
• Publication Date: 2024-10-30
• Authors: Zhiding Liu; Jiqian Yang; Qingyang Mao; Yuze Zhao; Mingyue Cheng; Zhi Li; Qi Liu; Enhong Chen
• Keywords: FOS: Computer and information sciences; Computer Science - Machine Learning; Machine Learning (cs.LG)

Multivariate time series forecasting plays a crucial role in various real-world applications. Significant efforts have been made to integrate advanced network architectures and training strategies that enhance the capture of temporal dependencies, thereby improving accuracy. On other hand, mainstream approaches typically utilize single unified model with simplistic channel-mixing embedding or cross-channel attention operations account for critical intricate inter-channel dependencies. Moreover, some methods even trade capacity robust prediction based on channel-independent assumption. Nonetheless, as data may display distinct evolving patterns due unique characteristics each channel (including multiple strong seasonalities trend changes), modeling could yield suboptimal results. To this end, we propose DisenTS, tailored framework disentangled general multivariate forecasting. The central idea DisenTS is potential diverse within decoupled manner. Technically, employs models, tasked uncovering pattern. guide learning process without supervision pattern partition, introduce novel Forecaster Aware Gate (FAG) module generates routing signals adaptively according both forecasters' states input series' characteristics. are derived from Linear Weight Approximation (LWA) strategy, which quantizes complex deep neural networks into compact matrices. Additionally, Similarity Constraint (SC) further proposed specialize an underlying by minimizing mutual information between representations.