Over-Smoothing

As the number of GNN layers increases, node features converge toward a stationary equilibrium point that depends only on the graph's global topology, not on the original input features. In a cellular interference graph, this means a congested picocell and an idle macrocell become indistinguishable after too many rounds of neighborhood aggregation, rendering the model incapable of localized resource allocation decisions.

The standard graph convolution operation acts as a low-pass filter on the graph's spectral domain, repeatedly smoothing node features along the graph Laplacian's eigenvectors. Each additional layer attenuates high-frequency components that encode fine-grained local differences between neighboring base stations, eventually retaining only the zero-frequency DC component that represents the global average feature vector across the entire cellular deployment.

A fundamental tension exists between model depth and feature discriminability:

• Shallow GNNs (1-3 layers): Preserve local node identity but fail to capture long-range interference dependencies across a metropolitan cellular grid
• Deep GNNs (4+ layers): Aggregate information from distant base stations but suffer exponential convergence of node representations toward a single cluster point
• Critical depth: The optimal number of layers is topology-dependent, often surprisingly shallow for dense urban deployments with high average node degree

Over-smoothing can be quantified by measuring the Dirichlet energy of node embeddings—the sum of squared differences between connected nodes' features. In a properly functioning GNN, this energy remains bounded away from zero. During over-smoothing, the Dirichlet energy decays exponentially with each additional layer, approaching zero as all connected base stations in the interference graph converge to identical representations regardless of their actual traffic load or spectral efficiency.

Residual connections that bypass one or more graph convolution layers inject the original node features directly into deeper representations, preserving local identity. In cellular topology models, this allows a GNN to simultaneously maintain a base station's unique hardware capability signature while still aggregating multi-hop interference context. Architectures like JK-Net (Jumping Knowledge Networks) explicitly concatenate outputs from all intermediate layers, letting the final classifier adaptively select the appropriate neighborhood scale for each node.

Over-smoothing catastrophically degrades performance on topology-dependent wireless optimization problems:

• Node classification: Congested vs. normal base station identification fails as all nodes converge to the same embedding
• Link prediction: Handover probability estimation between adjacent cells becomes random when node representations lose spatial distinctiveness
• Resource block allocation: The GNN cannot differentiate between high-interference and low-interference regions of the cellular graph, leading to uniform and suboptimal spectrum assignments

The foundational framework where over-smoothing occurs. In an MPNN, each node iteratively updates its state by aggregating 'messages' from its neighbors. With each layer, a node's receptive field expands exponentially, pulling in information from further across the graph. Over-smoothing is the direct result of this repeated averaging over increasingly overlapping neighborhoods, causing distinct node features to converge to a common equilibrium point.

The mathematical object that explains why over-smoothing happens. Defined as L = D - A (degree matrix minus adjacency matrix), its eigenvectors form the basis of the graph's Fourier domain. Repeated application of a low-pass graph convolution filter corresponds to multiplying the signal by powers of the Laplacian's eigenvalues. Since the smallest eigenvalues dominate, high-frequency (local) information is exponentially suppressed, leaving only the smooth, global components.

A related but distinct failure mode often confused with over-smoothing. Over-squashing occurs when information from an exponentially growing receptive field is compressed into a fixed-size node embedding vector, preventing long-range dependencies from being learned. While over-smoothing makes nodes indistinguishable, over-squashing makes distant nodes unable to influence each other. Both limit the effective depth of GNNs but require different solutions.

A partial mitigation strategy. GATs introduce a self-attention mechanism that dynamically weights the importance of neighboring nodes during aggregation. By learning to focus on specific, informative neighbors and ignore noisy ones, GATs can slow the homogenization process. However, attention alone does not prevent over-smoothing in very deep architectures; the underlying averaging dynamic still dominates without residual connections or normalization.

The frequency-domain perspective on the problem. Spectral convolution filters graph signals using the eigenvalues of the Graph Laplacian. Standard GCNs approximate a low-pass filter, which smooths signals across the graph. Over-smoothing is the extreme case where only the zero-frequency (DC) component survives. Understanding this spectral view motivates solutions like designing band-pass filters or using graph wavelets to preserve localized, high-frequency features.

The most common architectural defense. By adding a node's input features from layer l directly to its output at layer l+1, residual connections preserve a direct path for original, localized information. This prevents the initial node identity from being washed away by repeated neighborhood aggregation. Architectures like DeepGCN and JK-Net use skip connections to train GNNs with dozens of layers without catastrophic over-smoothing.