Few-shot sequence representation with supervised double-contrastive learning

Few-shot sequence representation, which is aimed at embedding sequentially few-shot data in a real space, is a foundational task in sequence pattern recognition. It is an intractable problem to effectively learn the intrinsic structural features within sequences with small sample sizes. In this paper, we propose to a supervised contrastive approach to represent small-sample data inspired by the state-of-the-art performance achieved by contrastive learning in unsupervised training of deep image models in recent years. Clusters of points belonging to the same class are pulled together in embedding space by supervised contrastive learning, while simultaneously pushing apart clusters of samples from different classes. An efficient method called supervised double-contrastive learning is proposed, which increases the sample size in disguise while supervised contrast to deal with small sample sequence data. The proposed method is experimentally evaluated on human activity recognition, protein recognition and disease recognition, and results obtained demonstrate its effectiveness and efficiency.