MSTVI: Multi-Scale Time-Variable Interaction for multivariate time series forecasting

doi:10.1016/j.knosys.2025.113551

MSTVI: Multi-Scale Time-Variable Interaction for multivariate time series forecasting

2025-06-01

0

PRE

AI

DOI：10.1016/j.knosys.2025.113551

Original

Original request for help

Save

Abstract

En 中文

Multivariate time series forecasting (MTSF) is widely used in economic forecasting, weather forecasting, public health monitoring and other fields. It involves multiple univariate series, and these series have complex interrelations at different scales. However, existing models lack the ability to capture multi-scale relationships among these variables, which limits the forecasting accuracy. Additionally, they typically treat time and variable modeling separately, making it difficult to fully grasp the deeper connections between them. To address these limitations, we propose a novel multi-scale framework called MSTVI, Multi-Scale Time-Variable Interaction model. MSTVI employs a pyramid-like hierarchical structure to efficiently capture and integrate features at different scales. At each level, a well-designed Time-Variable Interaction (TVI) module is employed to model complex relationships across both the time and variable dimensions simultaneously. The TVI module introduces a novel compression interaction mechanism to compress data along the time and variable dimensions, extract key information, and integrate it through a cross-interaction process to model time series, and it is generic enough to be plugged into an existing model to improve performance. MSTVI demonstrates superior performance across seven real-world benchmarks. Additionally, MSTVI's architecture ensures that both time and memory complexities scale linearly with the sequence length L and the number of variables M, achieving O(L) and O(M). Our code can be found in: https://github.com/liuquangao.

Keywords:

Time series forecasting

Multi-scale

Interaction

MLP

Real world datasets

MSTVI: Multi-Scale Time-Variable Interaction for multivariate time series forecasting

Abstract

Journal

Organization