Density-Guided Rank Correlation Graphs for Graph Convolutional Networks in Image Classification
Resumo
Graph Convolutional Networks (GCNs) have shown promising results in semi-supervised learning tasks, yet their effectiveness is highly dependent on the quality of the input graph. In image classification scenarios, graph construction remains a challenging step due to the general lack of inherent structural relationships between images. In this work, we propose a novel method to build input graphs for GCNs based on rank correlation measures between image similarity rankings. Each image is represented as a node, and edges are established according to the correlation between its ranked list of neighbors and those of other images. We introduce a density-guided strategy for automatically selecting the correlation threshold that controls the sparsity of the graph. Experiments conducted on three image classification datasets using three feature extractors and three GCN architectures show that the proposed correlation-based graphs outperform standard kNN and reciprocal kNN graphs in most cases, especially when used with the Simplified Graph Convolution (GCN-SGC) model. Our method surpasses several traditional and recent baselines, including techniques based on manifold learning and label propagation, while relying solely on contextual similarity through rank correlation without any post-processing refinement. The proposed approach's source code and documentation are publicly available at corgcn.lucasvalem.com.
