Graph Neural Network (GNN) for Interference

Unlike CNNs that assume a grid structure, GNNs inherently model the pairwise interference links between transceivers. This architectural prior encodes the physical reality that interference is a function of spatial proximity and channel overlap, not arbitrary pixel adjacency. The model learns that a node's state is fundamentally determined by its neighborhood aggregation, making it naturally suited to wireless topologies.

A critical property ensuring the GNN's output is independent of the arbitrary ordering of nodes in the input matrix. If you re-index the transceivers, the predicted interference pattern remains consistent. This is achieved through symmetric aggregation functions (sum, mean, max) in the message-passing layers, guaranteeing that the physical network topology, not the data structure's indexing, dictates the outcome.

The fundamental computational layer where nodes iteratively exchange information with their neighbors. Each node updates its hidden state by:

• Aggregating feature vectors from connected nodes (the interference sources)
• Combining this aggregate with its own current state via a learned function This process propagates information across the graph, allowing a node to infer the aggregate interference impact from both direct neighbors and multi-hop paths.

GNNs achieve scalability in dense networks by operating on local neighborhoods rather than the global adjacency matrix. The computational graph is defined by a node's k-hop neighborhood, meaning the model's complexity scales with the average node degree, not the total network size. This enables distributed execution where each radio can run inference based solely on its local interference graph, avoiding a centralized computational bottleneck.

The interference link is not just a binary connection; it carries rich physical attributes. GNN architectures encode edge features such as path loss, channel gain, distance, and frequency separation directly into the message computation. This allows the model to weigh messages based on the strength of the interference coupling, learning that a strong, close-proximity interferer should dominate the aggregation over a weak, distant one.

Wireless topologies are not static; nodes join, leave, and move. Advanced GNN architectures incorporate temporal message passing or recurrent units to handle evolving graph structures without full retraining. By processing a sequence of graph snapshots, the model learns to predict interference patterns under mobility, maintaining accurate resource allocation as the spatial relationships between transceivers continuously change.

A machine learning paradigm where multiple autonomous agents learn optimal policies through interaction and feedback within a shared environment. In spectrum sharing, each transceiver acts as an agent that learns to adjust its frequency and power to maximize throughput while minimizing interference. MARL is often combined with GNNs, where the graph structure defines the agent relationships and the GNN criticizes or coordinates the agents' actions. Key approaches include:

• Centralized Training, Decentralized Execution (CTDE): Agents share information during training but act independently during deployment
• Value Decomposition Networks: Learn to decompose a global reward signal into individual agent contributions

A stable state in a non-cooperative game where no individual player can gain an advantage by unilaterally changing their strategy. In interference management, a Nash Equilibrium represents a channel allocation where no single transmitter can improve its data rate by switching frequencies, given the choices of all other transmitters. GNNs can be trained to predict or converge to these equilibria by learning the underlying payoff structure of the interference game. The concept is critical for:

• Proving the stability of a learned allocation policy
• Modeling competitive spectrum sharing between rival network operators
• Designing distributed algorithms that guarantee convergence

A multi-dimensional, real-time geospatial database that integrates sensor data, propagation models, and regulatory policies into a comprehensive map of electromagnetic activity. A REM serves as the ground truth or input feature space for GNN-based interference models. The graph structure maps naturally to REM data:

• Nodes represent spatial grid points or active transmitters with features like received signal strength and modulation type
• Edges capture spatial adjacency or measured path loss between locations
• Graph convolutions propagate information across the map to estimate interference at unmonitored locations, enabling predictive spectrum allocation

A mathematical framework for solving coordination problems where multiple agents, each with local constraints, must agree on a globally optimal assignment of variables. In spectrum sharing, each radio has a local constraint on acceptable interference and must select a channel that satisfies a global objective. GNNs accelerate DCOP solving by:

• Learning to approximate the constraint graph's structure to prune the search space
• Predicting optimal variable assignments without running a complete distributed algorithm
• Generalizing across different network topologies, allowing a trained GNN to solve new DCOP instances without retraining

A calculated safety buffer representing the total allowable interference from all secondary users at an incumbent receiver. This margin ensures the incumbent's operational threshold is not exceeded. GNNs are uniquely suited to model aggregate interference because:

• The total interference at a receiver is the sum of contributions from all transmitters, a natural graph aggregation operation
• GNNs can learn non-linear interference interactions like intermodulation products that simple summation misses
• The model can predict when the aggregate margin is about to be exceeded and proactively trigger mitigation, such as reducing power or reassigning channels