TIFE: Tree-Structured Interactive Feature Enhancement for Multivariate Time Series Forecasting

In recent years, deep learning has emerged as one of the primary methods for multivariate time series (MTS) forecasting, achieving significant advancements across various domains. However, most deep learning models often fail to consider the multi-resolution characteristics of time series data, which may lead to information loss issues. In this paper, we explore the utilization of information from raw time series data at various resolutions and propose an effective feature extraction module, called Interactive Feature Enhancement (IFE). In IFE module, a three-branch structure is employed to extract the effective features from the raw MTS. Concretely, we first utilize the convolution operation to generate feature sequences and decompose these sequences and the raw time series into two sub-sequences(odd and even) by down-sampling separately. Next, the four sub-sequences are processed through convolutional layers, where the odd and even sub-sequences are interactively enhanced, respectively resulting in two feature-enhanced sub-sequences. Finally, we use convolutional interaction operations on these feature-enhanced sub-sequences to ensure data integrity. To fully exploit potential features in MTS, we propose a tree-structured architecture (named TIFE) by stacking the IFE module according to the properties of a complete binary tree. We conduct experiments on three MTS datasets of different domains and the experimental results show that TIFE has good forecasting performance. Compared to other baselines, TIFE significantly improves forecasting accuracy, further confirming its effectiveness and superiority.