ProSAR: Prototype-Guided Semantic Augmentation and Refinement for Time Series Contrastive Learning

ICLR 2026 Conference Submission11339 Authors

Keywords: Time series representation learning, contrastive learning, data augmentation

Abstract:

Contrastive learning has advanced the representation learning of vision, language, and graphs, yet its success hinges greatly on the data augmentation that helps preserve semantic contents while providing the view diversities. Multivariate time series, however, are noisy, non-stationary in nature, and largely opaque to the human inspection. Therefore, a direct use of the the hand-crafted transforms, such as jitter and scaling, may unfortunately destroy the critical temporal cues or introduce false negatives, weakening the performance of downstream tasks.
To address this, we propose ProSAR, a prototype-guided semantic augmentation and refinement framework for time series contrastive learning. Most critically, ProSAR's approach is founded on an information-theoretic principle for co-designing the semantic data augmentations and learnable prototypes, aiming to generate views that maximize the information about an associated semantic prototype while discarding the prototype-irrelevant content. ProSAR then implements this by introducing a novel prototype-conditioned semantic segment extraction mechanism, where the temporal characteristic segments are identified based on their dynamic time warping (DTW) alignment to these learnable time-domain prototypes, ensuring that the generated views can capture high-level semantic events. Building upon these temporal characteristic segments, the targeted augmentations, operating in both the time and frequency domains and informed by the DTW alignments, can thus preserve the temporal dynamics while constructing views that adhere to the information-theoretic objectives. Furthermore, prototypes are dynamically refined in a feedback loop, where the latent representations of these prototypes are refined via clustering under the prototypical contrastive training, and in turn guide evolution of the time-domain prototypes through a decoding consistency mechanism, thus fostering a progressive learning of robust representations. Experiments on diverse time-series benchmarks demonstrate that ProSAR outperforms recent contrastive and self-supervised representation learning methods in the downstream forecasting and classification tasks.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 11339