Clarify Confused Nodes Through Separated Learning

Graph neural networks (GNNs) have achieved remarkable advances ingraph-oriented tasks. However, real-world graphs invariably contain a certainproportion of heterophilous nodes, challenging the homophily assumption ofclassical GNNs and hindering their performance. Most existing studies continueto design generic models with shared weights between heterophilous andhomophilous nodes. Despite the incorporation of high-order messages ormulti-channel architectures, these efforts often fall short. A minority ofstudies attempt to train different node groups separately but suffer frominappropriate separation metrics and low efficiency. In this paper, we firstpropose a new metric, termed Neighborhood Confusion (NC), to facilitate a morereliable separation of nodes. We observe that node groups with different levelsof NC values exhibit certain differences in intra-group accuracy and visualizedembeddings. These pave the way for Neighborhood Confusion-guided GraphConvolutional Network (NCGCN), in which nodes are grouped by their NC valuesand accept intra-group weight sharing and message passing. Extensiveexperiments on both homophilous and heterophilous benchmarks demonstrate thatour framework can effectively separate nodes and yield significant performanceimprovement compared to the latest methods. The source code will be releasedsoon.