DABaCLT: A Data Augmentation Bias-Aware Contrastive Learning Framework for Time Series Representation

Contrastive learning, as an unsupervised technique, has emerged a prominent method in time series representation learning tasks, serving viable solution to the scarcity of annotated data. However, application data augmentation methods during training can distort distribution raw This discrepancy between representations learned from augmented contrastive and those obtained supervised results incomplete understanding information contained real trained encoder. We refer this bias (DAB), representing disparity two sets representations. To mitigate influence DAB, we propose DAB-aware framework for (DABaCLT). leverages features stream (RFS) extract data, which are then combined with create positive negative pairs learning. Additionally, introduce DAB-minimizing loss function (DABMinLoss) within contrasting module minimize DAB extracted temporal contextual features. Our proposed is evaluated on three classification including sleep staging (SSC) epilepsy seizure prediction (ESP) based EEG human activity recognition (HAR) sensors signals. The experimental demonstrate that our DABaCLT achieves strong performance self-supervised representation, 0.19% 22.95% accuracy improvement SSC, 2.96% 5.05% HAR, 1.00% 2.46% ESP, comparable approach. source code open-source.