Deep K-means Clustering Based on Generative Adversarial Framework
Yan Wang, Lei Pan · 2024
The deep clustering technology combines deep neural networks with clustering optimization to represent high-dimensional data as low-dimensional features that are more efficient for clustering, and can achieve good clustering performance in complex data with highly nonlinear structures. This paper proposes a deep K-means clustering model based on generative adversarial framework, which mainly includes three parts: an auto-encoder, a K-means clustering procedure, and a sampling layer. The low-dimensional features of the original data are firstly extracted by the auto-encoder, and then the K-means clustering is performed on them to obtain class labels and class centers. After that, the "true" samples and the "false" samples are obtained by means of the clustering results through the sampling layer. A discriminator is employed to correct and optimize the generator by discriminating the clustering results. Meanwhile, in order to enhance the inter-class separability of the model, we introduce two regularization terms into the discriminator loss function. The experimental results demonstrate that compared with the different kinds of clustering methods, the proposed method can effectively improve the clustering performance.