TS-MVP: Time-Series Representation Learning by Multi-view Prototypical Contrastive Learning

IoT and wearable devices generate large amounts of time series data daily, providing opportunities for the development of human-computer interaction and digital services through learning powerful representations from these rich data. While Masked Autoencoders (MAE) have been used for time series representation learning, contrastive learning has superior performance. However, existing contrastive learning methods often utilize perturbation operations that may disrupt the local and global structure of time series data, and they do not explicitly model the relationship between downstream classification tasks. In this paper, we propose a framework based on multi-view prototypical contrastive learning for learning multivariate time-series representations from unlabeled data. Our approach involves transforming the original data into time-based and feature-based views using innovative masking technology based on state transfer probabilities and then embedding them using an encoder along with the original data. Moreover, a novel prototype contrastive module is designed that learns similar outputs from different views using clustered soft labels generated by the original data and prototypes, which helps the model develop fine-grained representations that can be effectively integrated into classification tasks. We conducted experiments on four real-world time series datasets, and the results demonstrate that our proposed TS-MVP framework outperforms previous time series representation learning methods when training a linear classifier on top of the learned features.