Multi-feature based network for multivariate time series classification

Multivariate time series classification is widely available in several areas of real life and has attracted the attention of many researchers. In recent years, many multivariate time series classification methods have been proposed. However, existing multivariate time series classification methods focus only on local or global features and usually ignore the spatial dependency features among multiple variables. For this, we propose a multi-feature based network (MF-Net). First, MF-Net uses the global-local block to acquire local features through the attention-based mechanism. Next, the sparse self-attention mechanism captures global features. Finally, MF-Net integrates the local features and global features to capture the spatial dependency features using the spatial-local block. Therefore, we can mine the spatial dependency features of multivariate time series while incorporating both local and global features. We conducted experiments on UEA datasets and the experimental results showed that our method achieved performance competitive with that of state-of-the-art methods.