CO-COME: A Contrastive Label Disambiguation Framework with Combined MTS Feature Encoder for Partial-Label Multivariate Time Series Classification

ICLR 2026 Conference Submission15557 Authors

Keywords: Multivariate time series classification;partial label learning;

Abstract: Multivariate Time Series Classification(MTSC) is a conventional time series task applied in the fields of finance, healthcare, and weather forecasting. However, it is often plagued by a lack of high-quality labels in practical applications. To address the label quality issues in real-world scenarios, the Partial Label Learning(PLL) paradigm has been proposed. This paradigm solves the problem of ambiguous labels by allowing each training instance to be associated with a set of candidate labels. The superiority of PLL has been verified in the field of image classification. But due to the inherent difficulties in feature extraction for Multivariate Time Series(MTS) and the lack of appropriate data augmentation strategies, PLL has not been applied in MTSC tasks. Motivated by this, we propose a novel model: COntrastive label disambiguation framework with COmbined MTS feature Encoder(CO-COME), which integrates a contrastive learning-based label disambiguation framework with an efficient MTS feature representation encoder, CTFE. The contrastive learning module leverages label prototypes to effectively resolve label ambiguity under the PLL setting, meanwhile the CTFE encoder is designed to capture both explicit and latent representations of time series data, enabling robust and discriminative feature learning. Extensive experiments on 20 UEA benchmark datasets demonstrate that our model achieves state-of-the-art performance under partial-label conditions. Our method is available in https://github.com/Noname9971/CO-COME.

Primary Area: learning on time series and dynamical systems

Submission Number: 15557