Node Classification

Node classification typically operates in a semi-supervised setting where labels are only available for a small subset of nodes. The GNN leverages the graph topology to propagate label information from labeled nodes to unlabeled ones through iterative message passing.

• Training uses a masked loss function computed only on labeled nodes
• The model learns to generate embeddings that cluster by class in latent space
• Performance depends on the homophily assumption: connected nodes tend to share labels
• In cellular graphs, a few manually labeled base stations can enable classification across the entire network

Node classification can be transductive (predicting labels for unlabeled nodes within a fixed, known graph during training) or inductive (predicting labels for entirely new, unseen nodes or graphs).

• Transductive: The full graph structure is visible at training time; embeddings are learned specifically for that graph
• Inductive: Models like GraphSAGE learn a function that generalizes to new nodes, essential for dynamic cellular networks where user equipment constantly joins and leaves
• Inductive capability is critical for real-world deployment where the network topology evolves continuously

The predictive power of node classification comes from fusing two information sources: node features (intrinsic attributes) and graph topology (relational context).

• Node features for a base station: transmission power, current load, queue length, hardware type
• Topological signals: a node's position in the interference graph, its degree centrality, or the labels of its neighbors
• GNN layers learn to weight and combine these signals, often giving more weight to neighbors with similar features
• This fusion enables detecting a congested cell not just by its own metrics but by the state of its interfering neighbors

Several GNN architectures excel at node classification, each with distinct aggregation mechanisms:

• Graph Convolutional Network (GCN): Applies a first-order spectral convolution; simple and effective for homophilous graphs
• Graph Attention Network (GAT): Uses self-attention to learn dynamic importance weights for each neighbor, allowing the model to focus on the most relevant connections
• GraphSAGE: Samples and aggregates features from a fixed-size neighborhood, enabling inductive learning on massive graphs
• The choice depends on graph size, homophily ratio, and whether inductive capability is required

Deep GNNs for node classification face two fundamental challenges that degrade performance:

• Over-Smoothing: After too many message-passing layers, node representations become indistinguishable as each node's receptive field covers the entire graph. This destroys local discriminative information.
• Over-Squashing: Information from distant nodes is compressed into a fixed-size vector through bottlenecks in the graph, preventing the model from learning long-range dependencies.
• Mitigations include skip connections, graph rewiring, and limiting the number of convolutional layers to 2-4 for most cellular topologies

Node classification in cellular networks often faces severe class imbalance: normal operating states vastly outnumber failure or congestion events.

• Accuracy alone is misleading; use macro F1-score or weighted F1-score to account for class imbalance
• Precision and recall per class are critical: missing a congested cell (false negative) has higher operational cost than a false alarm
• ROC-AUC for binary classification measures ranking quality across thresholds
• In production, monitor per-class performance on a held-out test set that reflects the real-world label distribution

The foundational framework that formalizes how node classification operates in GNNs. Each node iteratively updates its hidden state by receiving and aggregating messages from its immediate neighbors. The message function computes information to transmit, the aggregation function (sum, mean, max) combines incoming messages, and the update function produces the new node representation. After multiple rounds of message passing, a node's embedding captures information from its extended neighborhood, enabling accurate label prediction.

A GNN architecture that introduces a self-attention mechanism to dynamically weigh the importance of neighboring nodes during aggregation. Instead of treating all neighbors equally, GAT computes attention coefficients that allow the model to focus on the most relevant connections for classification. In cellular networks, this enables a base station to prioritize signals from high-interference neighbors over distant, low-impact ones. Multi-head attention further stabilizes learning by running multiple attention mechanisms in parallel.

A critical failure mode in deep GNNs where node representations become indistinguishable after too many message-passing layers. As the receptive field expands exponentially, node embeddings converge to a similar value, destroying the discriminative power needed for classification. In cellular topologies, this means a congested base station and a lightly loaded one would appear identical to the classifier. Mitigation strategies include skip connections, layer normalization, and limiting network depth to 2-4 layers.

An inductive GNN framework that generates node embeddings by sampling and aggregating features from a node's local neighborhood. Unlike transductive methods that require the full graph at training time, GraphSAGE learns aggregator functions (mean, LSTM, pooling) that generalize to previously unseen nodes. This is critical for cellular networks where new user equipment constantly joins the topology. The sampling strategy also controls computational cost, making it scalable to massive, dynamic graphs.

A graph structure containing multiple types of nodes and edges, each with distinct feature spaces and semantic meanings. In cellular networks, a heterogeneous graph might include node types for base stations (features: transmission power, load), user equipment (features: signal strength, mobility), and edge servers (features: compute capacity). Node classification on heterogeneous graphs requires specialized architectures like RGCN or HGT that learn type-specific transformations and message-passing rules.

A technique for injecting information about a node's absolute or relative position within the graph structure into its initial features. Since GNNs are permutation invariant by design, they lack inherent awareness of a node's structural role. Positional encodings—derived from Laplacian eigenvectors, random walk probabilities, or shortest-path distances—provide this topological context. For node classification, this helps distinguish nodes with similar local neighborhoods but different global positions in the cellular topology.