<p>Graph convolutional network has been widely used in multi-view learning. However, most studies only consider the convolution process for individual view separately, ignoring the coupling interactions between different views. Apparently, the resulting under-utilization of consistency and complementarity in multi-view data will inevitably weaken the model’s performance. This work tries to solve above issue. First, by utilizing the granular-ball computing, we propose a novel topology construction method to construct paired-node connections for fully characterizing the inter-node consistency; Second, we design an ordered cross-view convolution framework to fully utilize inter-view complementarity. Specifically, granular-ball computing can achieve high-quality clusters. Then the topology characterizing inter-node consistency can be constructed according to the clusters. This topology construction method can provide more reliable connections for subsequent graph convolutional operations. In the ordered cross-view convolution framework, the view with the most features will be performed convolution operation earliest. The convolution results of each layer will participate into the convolution process of next layer in the next view with the second most features. The analogous operations will continue until the convolution process of the view with the least features is completed. The cross-view convolution can fully exploit the inter-view complementarity. Finally, the semi-supervised classification experiments show that our approach significantly outperforms existing multi-view methods, validate the importance of inter-node consistency and inter-view complementarity to multi-view learning.</p>