Regression-Based Clustering Network via Combining Prior Information

Abstract Despite the promising performance, existing regression-based clustering methods still have the following limitations. (1) They only extract the shallow discriminant features, resulting in unstable clustering performance on data with complex underlying subspaces. (2) It is difficult to optimize the objective due to the discretization of the elements in the cluster indicator matrix, resulting in suboptimal solution. (3) They fail to employ the structure prior information embedded in the clustering label matrix, resulting in suboptimal clustering performance. Targeting at above problems, we propose a novel Regression based Clustering network via Combining Prior Information (RC2PI), which consists of a convolutional auto-encoder, a priori information encoder, and a discriminator. Specifically, the auto-encoder is used to generate the ideal distribution to relax discrete cluster indicator matrix, which can help obtain optimal solution. The prior information encoder is employed to exploit the structure prior knowledge embedded in clustering label matrix, thereby boosting clustering via a self-supervised manner. The discriminator, as a connector of the above two sub-networks, is used for verifying the embedding process of prior information that will guide the auto-encoder to generate a more reliable actual distribution. Extensive experiments demonstrate the effectiveness of RC2PI over state-of-the-art methods.