Enhanced anchor contrastive multi-view representations learning network for clustering

Deep multi-view contrastive clustering has gained widespread attention due to its superior ability to classify samples and explore consistent information. However, existing models still suffer from the following limitations: 1) the high complexity of most models limits the ability of the model to deal with large-scale data; 2) the anchor points selected by the K-means algorithm may deviate from the true clustering distribution; 3) the instance-level contrastive learning (CL) tends to indiscriminately augment all the sample, which may generate false-negative pairs; 4) most methods directly rely on the embedding representations for label learning, which ignores the cluster-level label information and makes it difficult to obtain discriminative clustering structures. To address the above problems, we propose a novel Enhanced Anchor Contrastive Multi-view Representations Learning Network for Clustering (EACMVC). Specifically, we introduce the anchor representations in the model to alleviate the high complexity constraints. During the learning process of anchor representations, we design a weight factor to quickly focus on more representative anchor points and establish more reliable similarity relationships between anchor points and samples. To effectively eliminate the effect of false-negative anchor pairs due to misclassified boundary samples, we develop the global structure-guided anchor representations CL module (GSgARCL). Meanwhile, the anchor representations CL is more robust under the guidance of global structure. Furthermore, we develop a self-supervised label alignment module (SsLA) to maintain the label consistency. In turn, the aligned labels are used as supervised signals to further enhance the anchor representations. Finally, we learn the target distribution instead of obtaining the clustering results directly. Meanwhile, the Kullback-Leibler (KL) divergence is used to align target distribution with soft labels, which can learn a more accurate clustering structure. Each module of the proposed model is complementary and mutually supportive. Extensive experimental results demonstrate the efficacy of EACMVC compared with the state-of-the-art algorithms.