Multi-scale graph attention subspace clustering network

Deep subspace clustering (DSC) network which uses deep auto-encoders (DAE) mapping raw data into a latent space for clustering has achieved significant performance. However, when it comes to graph-based clustering, it fails to encode the node attribute and graph structure simultaneously, which degrades the clustering performance. Meanwhile, existing graph convolutional network (GCN) based clustering methods usually do not have a clustering-oriented objection function since the process of learning representation and clustering are separated. In addition, both aforementioned methods can neither catch the multi-scale information nor use the current clustering labels effectively, which lead to the suboptimal performance. To this end, we proposed a novel end-to-end framework called Multi-Scale Graph Attention Subspace Clustering Network (MSGA). By employing a novel GCN-based feature extraction module, it can effectively capture the node representation on graph-based datasets. Moreover, a multi-scale self-expression module is designed for obtaining a more discriminative coefficient representation from each layer of the encoder and a self-supervised module is introduced for supervising the learning of node representation. Specifically, we train and optimize them in a unified framework so that different modules can benefit from each other. Extensive experiments demonstrate the superiority of our method compared with several state-of-the-art methods.