Contrastive deep embedded clustering

Abstract

Deep embedded clustering is a popular unsupervised learning method owing to its outstanding performance in data-mining applications. However, existing methods ignore the difficulty in learning discriminative features via clustering due to the lack of supervision, which can be easily obtained in classification tasks. To alleviate this problem, we build a contrastive learning based deep embedded clustering method, i.e., CDEC. Specifically, our model adopts deep auto-encoders to learn a latent discriminative embedded clustering structure. To overcome the problem of lacking label information, the CDEC constructs positive samples and negative samples with the data reconstructed from the data itself and other data, respectively. By maximizing the distance between positive and negative ones, the CDEC can not only obtain the most representative features but also explore the discriminative features. Extensive experiments on several public datasets demonstrate that our method achieves the state-of-the-art clustering effectiveness. Our codes are available at: https://github.com/guoshuaiS/contrastive-deep.