Multi-scale graph diffusion convolutional network for multi-view learning

Multi-view learning has attracted considerable attention owing to its capability to learn more comprehensive representations. Although graph convolutional networks have achieved encouraging results in multi-view research, their limitation to considering only nearest neighbors results in the decrease on the ability to obtain high-order information. Many existing methods acquire high-order correlation by stacking multiple layers onto the model, yet they could lead to the issue of over-smoothing. In this paper, we propose a framework termed multi-scale graph diffusion convolutional network, which aims to gather comprehensive higher-order information without stacking multiple convolutional layers. Specifically, in order to better expand the receptive field of the node and reduce the parameter complexity, the proposed framework utilizes a contractive mapping to transform features from multiple views on decoupled propagation rules. Our framework introduces a multi-scale graph-based diffusion mechanism to adaptively extract the abundant high-order knowledge embedded within multi-scale graphs. Experiments show that the proposed method outperforms other state-of-the-art methods in terms of multi-view semi-supervised classification.