DeCITS: Incomplete time series clustering with deep contrastive learning

Time series data often contains missing values in real-world applications, making incomplete time series clustering a significant research focus in recent years. However, incorporating imputation and clustering within a model can lead to rapid error propagation from imputed values. Many existing time series clustering methods struggle to reduce the error propagation from imputation to clustering. Additionally, how to learn cluster-friendly representations from unlabeled time series data with temporal dynamics, and how to isolate the representation extracted from the clustering task without considering their interaction, remain challenging issues. In light of this, this study presents a Deep Contrastive Incomplete Time Series Clustering (DeCITS) method, which utilizes a unified framework for learning representations of incomplete time series in a contrastive learning manner. To mitigate error propagation, an adversarial strategy is incorporated to distinguish real and imputed values, encouraging the model to impute more realistic missing data. Consequently, DeCITS obtains cluster-friendly representations and achieves highly competitive clustering performance. Extensive experiments on six real-world datasets and eight UCR datasets demonstrate the superiority of DeCITS over existing state-of-the-art methods.