DCT-Based Channel Attention for Multivariate Time Series Classification

Abstract

This article introduces a novel DCT-based channel attention (DCA) mechanism for time series classification (TSC) using convolutional neural networks (CNNs). Traditional squeeze-and-excitation (SE) mechanisms rely on global average pooling to model channel-wise interdependencies, which may oversimplify complex temporal dynamics. The proposed DCA model leverages discrete cosine transform (DCT) coefficients to incorporate frequency-domain information, capturing a broader spectrum of temporal features. Two selection criteria are employed to identify the most informative DCT coefficients for constructing the attention map. The first criterion utilizes the lowest frequency coefficients, whereas the second criterion selects the coefficients exhibiting the highest energy to construct the attention map. Comprehensive experiments on twelve diverse TSC datasets demonstrate that DCA consistently outperforms state-of-the-art attention mechanisms, achieving an average improvement of $\text{2.2}{\%}$ in classification accuracy.