Attention-based hierarchical denoised deep clustering network

Clustering is a basic task of data analysis and decision making. Recently, graph convolution network (GCN) based deep clustering frameworks have produced the state-of-the-art performance. However, the traditional GCN has not fully learnt the structural information of the neighbors. Therefore, in this paper, we propose an attention-based hierarchical denoised deep clustering (AHDDC) algorithm to solve the problem, which enables GCN to learn multiple layers of hidden information and uses the attention mechanism to strengthen the information. Besides, we use a denoising autoencoder to reduce the influence of the data noise on the clustering. In AHDDC, Firstly, we input the feature vector of the original data into a denoising autoencoder (DAE) to learn the hidden representation; secondly, the representation information of the autoencoder and the structure information constructed by the KNN graph are passed into a hierarchical attentional graph convolutional network; finally, a self-supervision module is used to optimize the clustering results. Experimental results show the superiorities of our method over most advanced algorithms. Besides, the effectiveness of the proposed hierarchical, attention based and denoising improving strategies are also verified experimentally.