Contrastive learning for time series has a core problem: the method must decide which inputs are similar and which are dissimilar. In a new arXiv paper, the authors introduce TEX-MEX, described as a similarity-guided temporal contrastive learning approach. The abstract argues that this decision is especially hard for time series because few augmentations reliably preserve meaning.
Prior methods, the paper notes, have relied on two strategies, though the abstract in the source is truncated before those strategies are fully described. What is clear from the announcement is that TEX-MEX is positioned as a response to the augmentation problem, rather than as another fixed augmentation scheme.
The source is a single arXiv abstract, so no experimental results or comparisons with existing methods are available in the announcement. The paper's significance, as stated, lies in reframing contrastive learning for time series around similarity guidance rather than assuming that generic augmentations will work.