Multi-view Graph Convolutional Network with Fully Leveraging Consistency via Granular-ball-based Topology Construction, Feature Enhancement and Interactive Fusion

A research team has introduced MGCN-FLC (Multi-view Graph Convolutional Network with Fully Leveraging Consistency), a novel architecture designed to overcome fundamental limitations in how AI models process multi-view data. Traditional graph convolutional networks (GCNs) for multi-view learning often rely on K-nearest neighbors (KNN) for topology construction, where choosing the 'k' value artificially constrains the model. They also typically overlook consistency between features within a single view and fuse information from different views only after processing them separately, missing opportunities for deeper integration.

MGCN-FLC tackles these issues with three specialized modules. First, a granular-ball-based topology construction module clusters nodes into groups with high internal similarity, capturing inter-node consistency more naturally than KNN. Second, a feature enhancement module explicitly models relationships between different features within each view. Third, an interactive fusion module allows all views to communicate and influence each other during the learning process, not just at the end. This holistic approach to leveraging 'consistency'—in nodes, features, and views—led to the model outperforming current state-of-the-art methods in semi-supervised node classification, as validated on nine different datasets.