Exploring Depth in Graph Convolutional Networks (GCN): Architecture, Experiments, and Future Work

Graph Convolutional Networks (GCN) are needed for non‑grid data such as social or citation graphs, where traditional CNNs cannot be directly applied. The article begins with a brief introduction to GCN and the motivation for exploring deeper GCN models.

Unlike CNNs, which can be stacked to hundreds of layers (e.g., ResNet, DenseNet), typical GCNs are limited to 3–4 layers because deeper GCNs tend to diverge or over‑smooth node features. The article identifies three main reasons for this limitation: over‑fitting on insufficient data, feature over‑smoothing, and gradient vanishing.

To address these issues, three architectural enhancements are proposed:

Introduce residual connections (ResGCN) inspired by ResNet to mitigate gradient loss.

Adopt dense connections (DenseGCN) similar to DenseNet, concatenating input and output features to preserve gradient flow.

Apply a skip‑neighbor (dilated) convolution strategy to enlarge the receptive field without increasing parameter count.

The resulting deep GCN framework consists of three parts: the core GCN backbone, a feature‑fusion module (combining ResGCN and DenseGCN logic), and an MLP classifier that outputs node labels in an end‑to‑end training pipeline.

Experiments were conducted with a 28‑layer ResGCN. Results show that, unlike plain GCNs which diverge when deepened, the enhanced ResGCN maintains convergence and achieves higher accuracy as depth increases. Visual comparisons illustrate the widening performance gap between deepened and shallow GCNs.

The study concludes that adding skip‑connections and dilated convolutions enables deeper GCNs to train more easily and achieve better receptive fields, leading to consistent performance gains.

Future work includes exploring additional CNN‑inspired optimizations such as pooling, normalization, and advanced feature‑fusion techniques, testing alternative distance metrics for k‑NN graph construction, and refining kernel size and dilation rate selection.

Several GCN‑related tools used during development are listed for reference.

—END—