Contrastive Learning for Multivariate Time Series Classification: an Early Fusion Approach

In recent years, the use of contrastive learning methods for time series classification has shown promising results. State-of-the-art approaches leverage self-supervised contrastive learning representations that encode the underlying series’ features. On top of the latent representation, classification is addressed as a downstream task. While the effectiveness of contrastive models in univariate series classification is established, their portability towards a multivariate scenario is still debated. In this work, we explore the use of contrastive time series representations on multivariate data. We compare the performance of state-of-the-art contrastive models on 30 benchmark datasets and explore the use of an early fusion network to combine the input dimensions. The preliminary results show that incorporating a preliminary stage of information fusion is beneficial to improve the performance of state-of-the-art contrastive time series classifiers.