Graph Neural Network (GNN)

The foundational mechanism where nodes iteratively aggregate feature information from their neighbors to learn contextual embeddings. In spectrum graphs, a channel node receives updates from adjacent channels experiencing correlated interference.

• Aggregation function: Sum, mean, or max pooling of neighbor states
• Update function: Neural network layer combining aggregated message with current node state
• Iterations: Typically 2-4 hops balance expressiveness and oversmoothing
• Edge weights: Can encode physical distance or frequency separation between channels

Transforms each spectrum channel or cognitive radio into a dense, low-dimensional vector capturing its latent characteristics and relational context. These embeddings serve as input to downstream prediction tasks.

• Input features: Channel frequency, bandwidth, historical occupancy, SNR
• Embedding dimension: Typically 64-256 for spectrum applications
• Positional encoding: Added to distinguish nodes with identical features
• Output: Used for link prediction, node classification, or clustering of similar channels

A dedicated neural network module that computes the probability or strength of interference between two channel nodes based on their learned embeddings. This enables the GNN to forecast which frequency pairs will experience harmful co-channel or adjacent-channel interference.

• Input: Concatenated or dot-product of source and target node embeddings
• Architecture: Multi-layer perceptron with sigmoid output for binary interference classification
• Training signal: Historical interference measurements or spectrum monitoring data
• Application: Proactive channel selection to avoid predicted interference edges

The process of building a dynamic graph where edges evolve over time to capture both spatial proximity and temporal correlation between spectrum nodes. This graph structure is the critical input that determines what the GNN can learn.

• Spatial edges: Connect channels within same geographic area or frequency band
• Temporal edges: Link the same channel across consecutive time steps
• Correlation edges: Connect channels exhibiting similar occupancy patterns
• Dynamic re-weighting: Edge weights updated based on recent spectrum measurements

An enhancement to standard message passing where the GNN learns to assign different importance weights to each neighbor during aggregation. Graph Attention Networks (GATs) allow the model to focus on the most relevant interfering sources.

• Attention coefficients: Learned per edge, indicating neighbor relevance
• Multi-head attention: Multiple parallel attention mechanisms for stability
• Application: Prioritizing strong interferers over distant, negligible sources
• Benefit: Improved prediction accuracy in heterogeneous spectrum environments

A readout operation that aggregates all node embeddings into a single graph-level representation. This enables predictions about the entire spectrum environment, such as overall congestion level or total interference power.

• Global mean/max pooling: Simple but effective aggregation
• Hierarchical pooling: DiffPool or Top-K pooling for large graphs
• Set2Set: Learned permutation-invariant readout using attention
• Output: Single vector used for regression or classification of global spectrum state

The foundational framework unifying most GNN architectures. During each layer, nodes aggregate feature vectors from their neighbors, apply a differentiable permutation-invariant function (sum, mean, max), and update their own embedding via a learnable transformation. In spectrum graphs, this allows a channel node to incorporate occupancy states and interference levels from adjacent frequency bands, capturing spatial-spectral correlations without requiring a fixed grid structure.

A spectral-based GNN variant that performs convolution operations directly on graph-structured data. The layer-wise propagation rule uses a renormalized adjacency matrix with self-loops to aggregate features from a node's local neighborhood. For spectrum mobility, a GCN can process a graph where edges represent co-channel or adjacent-channel interference relationships, learning node embeddings that predict which frequency will experience a primary user return next based on correlated activity patterns across the topology.

A GNN architecture that introduces self-attention mechanisms to weight the importance of neighboring nodes during aggregation. Instead of treating all neighbors equally, GAT computes attention coefficients that allow a channel node to focus on the most relevant adjacent frequencies. This is critical in heterogeneous spectrum environments where a cognitive radio must dynamically prioritize information from strongly correlated channels while ignoring weak or noisy interference relationships.

An extension that processes dynamic graphs evolving over time by combining GNN layers with recurrent units (GRU/LSTM) or temporal convolutions. The spatial component captures instantaneous channel correlations, while the temporal component models how these relationships shift as primary users activate or deactivate. This architecture directly addresses the core challenge of spectrum mobility prediction: forecasting future edge-level interference based on historical spatio-temporal patterns in the spectrum graph.

An inductive framework that generates embeddings by sampling and aggregating features from a node's local neighborhood, enabling generalization to previously unseen nodes. Unlike transductive methods requiring full graph retraining, GraphSAGE allows a cognitive radio entering a new environment to immediately generate embeddings for unfamiliar frequency channels. The architecture supports multiple aggregator functions—mean, LSTM, pooling—providing flexibility for different spectrum sensing topologies.

The dual objectives of GNN training in spectrum applications. Node embedding maps each frequency channel to a dense vector capturing its occupancy characteristics and topological role. Link prediction uses these embeddings to estimate the probability of interference between two channels, even if never directly observed. This enables proactive spectrum handoff by predicting which idle channels are most likely to experience hidden node interference from returning primary users in adjacent bands.