Ranking Time-Frequency Contrastive Learning for Multivariate Time Series Classification

For multivariate time series classification, current research predominantly focuses on contrastive learning to acquire suitable representations. Despite their successes in enhancing accuracy and reducing label dependency, existing methods primarily concentrate on time-domain features, potentially neglecting the frequency-domain information inherent in time series. Additionally, the challenging problem of addressing the impact of false negative samples in contrastive learning remains unresolved. To tackle this, we propose a cross-modal architecture based on ranking time-frequency contrastive learning. This novel approach considers time-frequency consistency of time series, introducing an innovative time-frequency-based ranking loss to regulate the proximity between positive/negative samples and the anchor point, thereby mitigating issues related to false negatives. Extensive experiments validate the effectiveness of our proposed method in advancing multivariate time series classification