Graph t-SNE multi-view autoencoder for joint clustering and completion of incomplete multi-view data

In recent years, incomplete multi-view clustering has attracted much attention and achieved promising performances through the use of deep learning. However, only a few prior methods are concerned with joint missing data recovery and clustering. In this paper, we present a graph t-SNE multi-view autoencoder (GTSNE-MvAE) for this task. We formulate the view completion problem as a multi-view multivariate regression and reconsider the autoencoder for this task. First, a multi-view encoder augmented with graph-convolutional layers and the t-SNE regularization loss extracts unified representations from incomplete multi-view features. Then, the representations are fed into view-specific decoders to regress the features of views, through which the missing views are recovered. Notably, the unified representations learned by our model are cluster-friendly. Our simple method achieves competitive clustering performances on 9 challenging public benchmarks while keeping a stable training process and hyperparameter insensitivity.