Soft Contrastive Learning for Irregular Multivariate Time Series

Irregular multivariate time series (IMTS) are characterized by irregular observation times, resulting in 1) misaligned time points across features (i.e., misalignment) and 2) inconsistent intervals between observations (i.e., inconsistency). However, existing time series methods often overlook these irregularities, leading to suboptimal performance, or depend on large labeled datasets. To this end, we introduce SITS, a simple yet effective soft contrastive learning strategy tailored for IMTS, where pairs are constructed from a single instance that shares the same irregularities, rather than from different instances with varying irregularities. Specifically, different views of a single instance are generated with varying masking ratios, where higher masking ratios correspond to smaller soft label values. Furthermore, we propose SeqTAND, a model architecture that handles misalignment and inconsistency in a sequential manner, which is shown to be more effective than addressing them in parallel. Experimental results demonstrate that SITS outperforms state-of-the-art methods in both classification and interpolation tasks.