Graph-Level Regression

A fundamental mathematical property ensuring that the predicted scalar output for a graph remains unchanged regardless of the arbitrary ordering of input nodes. This is non-negotiable for graph-level regression on cellular topologies.

• Why It Matters: A network's aggregate spectral efficiency must be identical whether base station A is indexed as node 1 or node 50.
• Enforcement: Achieved through symmetric aggregation functions like sum, mean, or max in the readout layer.
• Isomorphism: Two structurally identical graphs with different node orderings must produce the exact same regression value.

The critical architectural component that collapses all learned node embeddings into a single fixed-size vector representing the entire graph, from which the final scalar value is regressed.

• Global Sum Pooling: Sums all node features; sensitive to graph size, useful for predicting total network throughput.
• Global Mean Pooling: Averages node features; provides a normalized view, ideal for average spectral efficiency.
• Set2Set & Attention Pooling: Learnable, weighted aggregations that allow the model to focus on the most relevant nodes (e.g., congested cells) for the prediction task.

A critical failure mode in deep GNNs where node representations become indistinguishable after too many layers of aggregation. This directly sabotages graph-level regression by destroying the local information needed for an accurate global prediction.

• Cause: Repeated Laplacian smoothing acts as a low-pass filter, averaging away discriminative features.
• Impact on Regression: The model loses the ability to distinguish between a high-interference and low-interference network topology.
• Mitigations: Use skip connections (ResGNNs), Graph Attention Networks (GATs) to dynamically re-weight edges, or limit model depth.

A theoretically grounded GNN architecture proven to be maximally expressive for graph-level tasks, matching the discriminative power of the Weisfeiler-Lehman (WL) graph isomorphism test.

• Expressiveness: GIN can distinguish between graph structures that simpler GNNs like GCNs or GraphSAGE cannot, leading to more accurate regression on subtly different network topologies.
• Aggregation Formula: Uses a sum aggregator combined with a multi-layer perceptron (MLP) to model an injective function over the multiset of neighbor features.
• Use Case: Critical when predicting energy efficiency requires differentiating between network snapshots with identical node degree distributions but different interference patterns.

The standard evaluation metrics for graph-level regression, quantifying the difference between the predicted scalar value and the ground truth.

• MAE (L1 Loss): Measures the average absolute difference. Less sensitive to outliers, providing a linear penalty for errors in predicting metrics like average user throughput.
• RMSE (L2 Loss): Penalizes large errors more heavily due to squaring the difference. Preferred when large prediction misses for peak energy consumption are disproportionately costly.
• Training Objective: The GNN's weights are optimized via backpropagation to directly minimize one of these loss functions on a dataset of graph-scalar pairs.