Multivariate time series classification with dual attention network

TL;DR

This paper introduces DA-Net, a dual attention network that effectively captures both local and global features in multivariate time series classification by combining SEWA and SSAW layers.

Contribution

It presents a novel dual attention network architecture that integrates local and global feature extraction for improved multivariate time series classification.

Findings

DA-Net effectively captures local sequence fragments.

DA-Net models long-range dependencies.

The approach improves classification performance.

Abstract

One of the topics in machine learning that is becoming more and more relevant is multivariate time series classification. Current techniques concentrate on identifying the local important sequence segments or establishing the global long-range dependencies. They frequently disregard the merged data from both global and local features, though. Using dual attention, we explore a novel network (DA-Net) in this research to extract local and global features for multivariate time series classification. The two distinct layers that make up DA-Net are the Squeeze-Excitation Window Attention (SEWA) layer and the Sparse Self-Attention within Windows (SSAW) layer. DA- Net can mine essential local sequence fragments that are necessary for establishing global long-range dependencies based on the two expanded layers.