Divide and Contrast: Learning Robust Temporal Features without Augmentation

Researchers introduced Di-COT, a self-supervised time-series representation method that learns temporal features without data augmentation.

Di-COT splits each window into overlapping sub-blocks and contrasts them, avoiding repeated encoder passes while preserving useful temporal structure.

Across six datasets and standard UCR/UEA benchmarks, it delivered state-of-the-art accuracy and faster training.

The method improves efficiency for classification, clustering, and cross-dataset transfer by learning more transferable features at lower cost.