A Label-Refinement Stage Before GCN Training Beats the Baseline by 1.67 Points Under Label Noise

arXiv 2609.22053·low signalXBlueskyLinkedInCopy link

Graph Convolutional Networks propagate corrupted supervision through the graph, so PCC+GCN inserts Particle Competition and Cooperation as a label-refinement pass before training: particle domination dynamics flag suspicious labeled nodes and decide whether each label is preserved, removed or reassigned, optionally with feature-based kNN edges added to the PCC graph while the GCN still trains on the original structure. Across ten datasets from the NoisyGL benchmark under Uniform, Pair and Random noise, it takes the highest overall average accuracy and best average rank, averaging 1.67 percentage points over the baseline GCN across clean and noisy settings. Under instance-dependent noise it stays competitive with the best robust methods at substantially lower execution time.