Incomplete multi-view classification via graph neural network on heterogeneous graph

Incomplete multi-view classification aims to classify the multi-view data with missing views. Several works have been proposed to impute the missing views and then conduct the existing multi-view classification, or conduct these two tasks simultaneously. However, the final classification performance of these works depends heavily on the missing view imputation. Unlike these existing works, in this paper, we propose a novel Incomplete Multi-view Classification method with Graph neural network on Heterogeneous graph (IMCGH). We transform the multi-view data into a heterogeneous graph by mapping each sample in each view to a node of a different type. Missing views can be regarded as learning using a subgraph of the heterogeneous graph, allowing our method to conduct incomplete multi-view classification naturally. We also design the loss functions based on mutual information to exploit the consistency and complementarity of information within multi-view data. Experimental results on several benchmark datasets illustrate the effectiveness and superiority of the proposed method compared with its state-of-the-art competitors in the transductive and inductive learning tasks.