Deep Clustering of Time Series Based on Neighborhood Structure Learning

Deep learning has shown remarkable potential in enhancing time series clustering by leveraging its powerful data representation capabilities. However, most deep learning methods struggle to simultaneously optimize feature extraction and clustering objectives, hindering the ability of the clustering task to effectively guide the feature extraction process. In addition, existing approaches for time series often fail to account for the learning of neighborhood structures within the data, thereby limiting their ability to capture the intrinsic relationships among data points. In this paper, we propose a deep clustering method for time series based on neighborhood structure learning, designed to simultaneously optimize feature extraction and clustering objectives. The proposed model incorporates an attention-based neighborhood structure fusion module, which integrates features from corresponding layers of the autoencoder and the graph convolutional network. Additionally, to enhance the robustness of clustering, an adversarial regularization module is introduced to improve the stability of graph representations. Extensive experiments conducted on multiple time series datasets demonstrate the superior performance of our proposed model.