Abstract | ||
---|---|---|
Multivariate time series (MTS) classification is a challenging problem due to the complex nature of data, especially for tasks with spatial dependencies such as three-dimensional (3D) skeleton sequences, traffic flow sequences, sign language sequences, and activity acceleration sequences. Existing methods do not fully exploit spatial features in these MTS, so the major challenge is how to effectively extract these spatial features and integrate them into an end-to-end network. Therefore, we propose the Spatio-Temporal Dependence Learning Network (STDL-Net), which is an attention-based network for MTS with explicit spatial dependencies. The STDL-Net first simultaneously extracts spatial dependencies and models short-term temporal dependencies via a sparse convolutional neural network (CNN) to obtain spatio-temporal dependence units (STDUs). Then the STDL-Net leverages a long short-term memory (LSTM) network with an attention mechanism to model long-term temporal dependencies and adaptively select the most discriminative STDUs for MTS classification tasks. In this way, the STDL-Net acts as an end-to-end model which captures discriminative sample-specific spatial features at each time step and models underlying multi-scale dynamics in the MTS data. Experiments on 12 MTS benchmark datasets with explicit spatial dependencies and 3 skeleton-based action recognition tasks are conducted. The proposed STDL-Net yields better performance than existing methods and the visualized analysis demonstrates the effectiveness of learning STDUs. |
Year | DOI | Venue |
---|---|---|
2019 | 10.1016/j.ins.2019.07.007 | Information Sciences |
Keywords | Field | DocType |
Spatio-temporal dependencies,Attention mechanism,Multivariate time series classification,Convolutional neural network,Long short-term memory | Traffic flow,Convolutional neural network,Multivariate statistics,Exploit,Sign language,Artificial intelligence,Acceleration,Discriminative model,Machine learning,Mathematics,Learning network | Journal |
Volume | ISSN | Citations |
503 | 0020-0255 | 1 |
PageRank | References | Authors |
0.35 | 0 | 5 |
Name | Order | Citations | PageRank |
---|---|---|---|
Qianli Ma | 1 | 20 | 5.80 |
Shuai Tian | 2 | 3 | 1.09 |
jia wei | 3 | 4 | 3.09 |
Jiabing Wang | 4 | 5 | 2.43 |
Wing W. Y. Ng | 5 | 528 | 56.12 |