Diffusion-Guided graph generation for multi-view semi-Supervised classification.

Abstract

A considerable number of research focus on multi-view learning because of its ability to obtain better representations. Although many variants of graph convolution have been well adapted to multi-view semi-supervised tasks, most of them are limited by the fact that multi-view datasets often lack a natural graph structure. In addition, existing methods directly utilize original feature space to acquire graph structure without considering the possible presence of noise and synergistic acquisition of representation learning. To tackle the aforementioned challenges, this paper proposes a graph generation deep learning method, called multi-view diffusion-guided graph generation, consisting of two modules: diffusion module and aggregation module. The former utilizes a heat diffusion equation to act as a low-pass filter, obtaining a purified feature space from the original features for each view. The latter retains the original information with residual connection and generates a consistent feature representation space and its adjacency matrix. Accordingly, these outputs are utilized to calibrate parameters across several graph backbone networks for each view. Finally, a fusion method is employed to consolidate information from the above multiple graph backbone networks. Extensive experimental results demonstrate the superior performance of our proposed method, compared to other state-of-the-art multi-view semi-supervised approaches.