Negative samples filter of contrastive learning for time series classification

As an unsupervised learning method, contrastive learning has achieved remarkable success in the field of computer vision. However, issues such as false negative samples and hard negative samples can significantly impair its effectiveness. Addressing these issues is therefore crucial for improving contrastive learning. While current research on handling these challenges mainly focuses on image data, there is limited exploration of contrastive learning for time series data. In this paper, we propose a negative samples filter in the embedding space to investigate the impact of hard negative samples on time series contrastive learning. We conducted extensive experiments on six different time series datasets to examine the effect of the negative samples filter on classification performance, both in unsupervised and supervised settings. The results demonstrate that in the unsupervised case, some of the most difficult samples can degrade classification performance, while in the supervised case, more difficult samples are beneficial for classification. Furthermore, we applied our filter function to other contrastive learning baselines for time series, achieving superior results compared to previous baselines, and outperforming other baselines that address false negative and hard negative samples.