Paper Title
LABEL CENTRIC PROTOTYPE LEARNING FOR GRAPH NEURAL NETWORKS

Abstract
Graph Neural Networks (GNNs) are used for node classification by learning about the node from neighbouring nodes. They usually work well when connected nodes belong to the same class as the information is passed to the neighbouring nodes, but in many cases, this is not true. When neighbouring nodes connected to different classes, thismakes the classification difficult for the model and performance drops.In this work, we tried to address this problem using prototype-based ideas. For homophilic graphs, we propose Global LocalCross-Prototype Repulsion (GL_CPR), which reduces the effect of connection of nodes between different classes which is usually less in homophilic graphs. For heterophilic graphs, we introduce Label-Centric Anchor Network (LCAN), where class prototypes act as stable anchor points for the nodes to improve classification.From our experiments, LCAN improves performance clearly on heterophilic datasets, while GL_CPR gives consistent improvements on homophilic graphs. Keywords - Global-Local Learning, Graph Neural Networks, Heterophily, Node Classification, Prototype Learning