Multivariate time series classification (MTSC), one of the most fundamental time series applications, has not only gained substantial research attentions but has also emerged in many real-life applications. Recently, using transformers to solve MTSC has been reported. However, current transformer-based methods take data points of individual timestamps as inputs (timestamp-level), which only capture the temporal dependencies, not the dependencies among variables. In this paper, we propose a novel method, called SVP-T. Specifically, we first propose to take time series subsequences, which can be from different variables and positions (time interval), as the inputs (shape-level). The temporal and variable dependencies are both handled by capturing the long- and short-term dependencies among shapes. Second, we propose a variable-position encoding layer (VP-layer) to utilize both the variable and position information of each shape. Third, we introduce a novel VP-based (Variable-Position) self-attention mechanism to allow the enhancing the attention weights of overlapping shapes. We evaluate our method on all UEA MTS datasets. SVP-T achieves the best accuracy rank when compared with several competitive state-of-the-art methods. Furthermore, we demonstrate the effectiveness of the VP-layer and the VP-based self-attention mechanism. Finally, we present one case study to interpret the result of SVP-T.

Time series classification (TSC) has been one of the most fundamental problems of time series data. Time series shapelets (or simply, shapelets) are discriminative subsequences that have been recently found both effective and interpretable for solving TSC. However, shapelet discovery is known to be computationally costly. Meanwhile, matrix profile has been recently proposed for efficient motif discovery and anomaly detection. Our preliminary experiment shows that a direct adoption of the matrix profile on TSC does not bring superior classification accuracy. We have identified two main issues of such an adoption: 1) discords as “shapelets”, and 2) lack of shapelet diversity. In response to these issues, we propose instance profile for shapelets, called IPS, for shapelet discovery for TSC. The main challenge is to utilize the instance profile (IP) to capture the characteristics of shapelets in a robust manner and then to discover high-quality shapelets efficiently. First, we use our IP to generate abundant shapelet candidates. We next efficiently prune candidates that do not align with the definition of shapelets using a novel distribution-aware bloom filter (DABF). Three utility functions are proposed to measure the shapelet candidates and DABF is used to efficiently compute the functions. We have conducted comprehensive experiments on IPS with 12 competitive state-of-the-art methods using UCR Archive datasets. The efficiency is on average 25 times faster than that of BSPCOVER (the current state-of-the-art method). The accuracy of IPS is comparable to or higher than that of existing work. Furthermore, we select one case study to illustrate the interpretability of the shapelets.