TimesMR: Unlocking the Potential of MLP and RNNs for Multivariate Time Series Forecasting

Huiqiang Wang, Jieming Shi, Li Qing

Keywords: Multivariate time series forecasting, MLP-based multiscale modeling, Grouped BiRNN, Effectiveness and efficiency.

TL;DR: Lightweight multiscale MLP to capture temporal dependency; Effective Grouped BiRNN structure to explore variable relationship

Abstract:

Multivariate time series forecasting is critical across various domains, requiring effective modeling of temporal dependencies and variable correlations. Existing multi-scale models, while effective for temporal dependency modeling, often rely on complex and computationally expensive feature extractors. Similarly, attention mechanisms, though powerful for capturing variable relationships, suffer from high complexity on high-dimensional datasets and introduce noise from weakly related variables, leading to performance degradation.
To address these challenges, we propose TimesMR, a novel model with two key innovations. First, we design multi-scale MLP modules, namely multi-patch MLP and multi-downsampling MLP, to enhance temporal dependency modeling with lightweight and efficient architectures. Second, we introduce grouped bidirectional RNNs for efficient variable correlation modeling, which reduce computational costs while preserving performance by grouping variables and capturing both intra- and inter-group correlations.
Extensive experiments on sixteen datasets demonstrate that TimesMR achieves state-of-the-art performance, surpassing eighteen existing models. Our contributions include novel plug-and-play modules for temporal and variable modeling, offering an effective and efficient solution for multivariate time series forecasting. The code is available at https://anonymous.4open.science/r/TimesMR.

Primary Area: learning on time series and dynamical systems

Submission Number: 9311