Inconsistency-Aware Graph Convolutional Networks for Multi-View Classification

Graph convolutional networks (GCNs) have attracted significant attention in the field of multi-view learning, as they effectively extract intricate information from diverse features. Although GCN-related studies have achieved encouraging performance, the majority rely heavily on the consistency principle to capture shared information from multiple views. Accordingly, the inconsistency principle, which is beneficial for exploring complementary information among multiple views is rarely involved. Therefore, we propose Inconsistency-Aware GCN (IAGCN), a novel framework that addresses the problem of the insufficient consideration of inter-view inconsistencies. Our approach incorporates inter-view inconsistency encoding, inter-view feature enhancing, and intra-view topology augmentation into a comprehensive framework. Specifically, we firstly design an inter-view inconsistency encoder to reveal the inconsistency between pairwise features. Then, we devise a feature enhancing module with an inconsistency perception to feed the complementary information into the intra-view features. Additionally, we employ a learnable intra-view topology augmentor to refine node relationships. We embed these three modules into the training process to capture the upgraded embeddings. Extensive semi-supervised classification experiments on six public datasets confirm that IAGCN significantly surpasses state-of-the-art methods.