Multivariate time-series classification using memory and attention for long and short-term dependence ^⋆

Time series classification (TSC) is one of the most challenging research topics in data mining, and can finds its wide applications in biomedical engineering and clinical prediction. In recent years, deep learning (DL) has shown impressive performance in TSC research due to its end-to-end capabilities. In contrast to univariate TSC (UTSC), multivariable TSC (MTSC) is more prevalent and its effectiveness depends on long and short-term dependencies to a certain degree. However, mainstream DL models mainly construct various complex frameworks to extract features, and with the over-fitting problem. This paper proposes a comprehensive DL structure, namely the Multivariable Multi-Scale Attention Gated Cycle Unit Fully Convolutional Network (MMAGRU-FCN), which can effectively address the long and short-term dependence in MTSC by integrating memory and attention mechanisms at multiple scales. The proposed model can achieve adaptive characteristic calibration and capture significant features simultaneously. Extensive experiments demonstrate the superior performance of our model over state-of-the-art DL networks in terms of faster convergence, better stability, and in the case that parameters are close to the minimum threshold of comparison. Moreover, the proposed model consistently achieves the highest classification accuracy across different time series lengths. Finally, we verify the performance of the proposed model for various classification scenarios.