TSCCL: A Contrastive Learning Framework for Time Series Clustering with Hurst Exponent

Time series clustering plays a vital role in various fields, including power grids and finance. In recent years, contrastive learning has been widely adopted to improve the performance of clustering algorithms. However, these methods were initially designed for image data. When applied to time series, which exhibit unique characteristics such as mean reversion, the labels of generated samples often do not match the labels of the original samples. This misalignment occurs because the nature of time series data is not sufficiently preserved during augmentation. Therefore, ensuring that the inherent properties of time series data are maintained during augmentation has become a critical issue. In this paper, we firstly introduce a novel data augmentation method called HTGAN (Hurst-Time series Generative Adversarial Network), which is integrated into a contrastive learning framework. Moreover, we propose a new clustering framework, TSCCL (Contrastive Learning Framework for Time Series Clustering), specifically designed for financial time series data clustering. Experimental results demonstrate that TSCCL outperforms all baseline methods, highlighting the effectiveness of HTGAN in financial time series data augmentation.