Multi-view Graph Convolutional Networks with Differentiable Node Selection

Multi-view data containing complementary and consensus information can facilitate representation learning by exploiting the intact integration of multi-view features. Because most objects in real world often have underlying connections, organizing as heterogeneous graphs is beneficial to extracting latent among different objects. Due powerful capability gather neighborhood nodes, this paper, we apply Graph Convolutional Network (GCN) cope with graph originating from data, which still under-explored field GCN. In order improve quality network topology alleviate interference noises yielded fusion, some methods undertake sorting operations before convolution procedure. These GCN-based generally sort select confident nodes for each vertex, such picking top- k according pre-defined confidence values. Nonetheless, problematic due non-differentiable operators inflexible embedding learning, may result blocked gradient computations undesired performance. To these issues, propose a joint framework dubbed Differentiable Node Selection (MGCN-DNS), constituted an adaptive fusion layer, module differentiable node selection schema. MGCN-DNS accepts multi-channel graph-structural inputs aims learn more robust through neural network. The effectiveness proposed method verified rigorous comparisons considerable state-of-the-art approaches terms semi-supervised classification tasks, experimental results indicate that achieves pleasurable performance on several benchmark datasets.