Time series augmentation possessing the dual strategy is essential in successfully applying deep models in time series classification. Mainly, it holds significant research and practical value in mitigating challenges such as data scarcity and class imbalance. For instance, the infrequent occurrence of abnormal incidents in industrial conditions makes capturing anomalous samples challenging. Pattern mixing has emerged as an advanced and promising branch in time series augmentation, whose core ideas adopt the feature similarity of the time series. Although existing approaches perform better in improving diversity, they will inevitably introduce features unrelated to classification decisions. The core issue is they adopt to mix the global region with similar features. It may lead to abnormal augmented samples due to excessive irrelevant features, interfering in classification decisions.
To solve the problems, a research team led by Pin LIU published their new research on 15 December 2024 in Frontiers of Computer Science co-published by Higher Education Press and Springer Nature.
This paper aims to improve the feature diversity of training samples while expanding the data scale by mixing the locally similar regions in the two intra-class time series, improving the classification performance. The core steps are as follows: Similarity-based local region matching, Dynamic selecting matched segment and Mixing matched segment. We perform experiments on ten representative datasets from the UCR2018, including the classification performance, the time consumption and the influence of batch size. The results are recorded in the Table1-2 and Figure 2. The experimental results show that the proposed method is better than others and consumes less time. In addition, no matter how big the Batch Size is during the training process, this method is still effective and improves classification performance.
DOI: 10.1007/s11704-024-40110-9