MGTC: Multi-Granularity Temporal Aware Time Series Classification

Self-supervised learning (SSL) can extract useful temporal representations for time series classification (TSC) tasks. However, existing methods with subsequence and instance-level augmentation lead to the loss of global information and inductive bias. Meanwhile, neglecting multi-granularity temporal representations also poses challenges for modeling complex temporal structures and relationships. Therefore, in this work, we propose a multi-granularity temporal-aware time series classification method, called MGTC, that learns multi-granularity temporal representations, enhancing the ability to perceive class discrepancies. Specifically, we propose a multi-density masking strategy to adapt the dynamic time-varying patterns for learning comprehensive temporal representations. Next, we employ cluster-wise constraint to hierarchically aggregate these representations at the instance level. Finally, we design two self-supervised tasks: i) granularity-aware contrastive learning, to extract intra-instance fine-grained temporal structures and inter-instance coarse-grained class relationships, and ii) cross-view prediction pretext task, to capture global contextual temporal dependencies. We conducted comprehensive experimental evaluations on various types of dataset, and the experimental results validated the effectiveness of our method. The code is publicly available at https://github.com/mrxiliang/MGTC Note to Practitioners—Time series classification (TSC) is widely used in many real-world applications, such as industrial fault diagnosis and healthcare monitoring. To guarantee classification accuracy, the TSC models must have high representation learning ability. However, many existing methods fail to capture global information and different granularities of temporal relations, hindering accurate classification. To address this challenge, this paper proposes a multi-granularity temporal-aware time series classification method (MGTC), capturing multi-granularity temporal representations from unlabeled data, enhancing the ability to perceive class discrepancies. The proposed multi-density masking strategy could adapt to the time-varying patterns for fully capturing temporal representations. The cluster-wise constraint could make the learned representations more compact in the embedding space. The granularity-aware contrastive learning, enable MGTC to extract fine-grained temporal structures and coarse-grained class relationships. The experimental results demonstrate the excellent performance of MGTC in TSC tasks.