Graph Convolutional Networks Based on Aggregated Multilevel kNN Graph for Image Classification

• Gustavo Rosseto Leticio UNESP
• Vinicius Atsushi Sato Kawai UNESP
• Lucas Pascotti Valem USP
• Daniel Carlos Guimarães Pedronette UNESP

Resumo

Graph Convolutional Networks (GCNs) rely heavily on the quality of the input graph, which is often defined by a fixed neighborhood size. This work explores two complementary strategies to make GCN-based classification more robust. First, we build multiple graphs with different neighborhood sizes to capture diverse structural information. Second, we leverage the full softmax output distributions to apply confidence-based weighting and combine predictions through lightweight ensembles. This approach improves accuracy without modifying the GCN architecture or training process. Results show consistent gains across different models and graph configurations, highlighting the benefit of integrating multi-scale graph structures and uncertainty signals in the decision process.

Palavras-chave: Training, Visualization, Uncertainty, Accuracy, Graph convolutional networks, Predictive models, Feature extraction, Vectors, Robustness, Manifold learning