The Time-Sequence Prediction via Temporal and Contextual Contrastive Representation Learning

The time series classification tasks have commonly faced the problem, i.e., lower labelled time series data and higher labelling costs. Regarding this issue, some researchers try introducing representation learning into the time series classification task. Moreover, recently the researcher proposed a model called TSTCC. TS -TCC combines transformer and representation learning and has achieved promising performance. Therefore, we will predict the Time-sequence via Temporal and Contextual Contrastive Representation Learning (PTS-TCC). PTS-TCC tends to perform better than TS-TCC in robustness. PTS-TCC consists of four modules: cluster module, data hidden representation learning module, temporal hidden representation learning module, and contextual hidden representation learning module. Extensive quantitative evaluations of the HAR (Human Activity Recognition), Epilepsy (Epilepsy Seizure Prediction) and Sleep-EDF (Sleep Stage Classification) datasets verify the effectiveness of our proposed PTS-TCC. In contrast to SOTA, the average accuracy rate of PTS-TCC improves by 5% in HAR, Epilepsy and Sleep-EDF.