Interference Graph Construction

Each vertex in the graph represents a distinct communication link, defined as a specific transmitter (Tx) and its intended receiver (Rx) pair. The vertex's feature vector typically includes the direct channel gain between the Tx and Rx, the current transmit power, and the link's quality-of-service (QoS) requirement. This node-centric view allows the GNN to learn a per-link policy for resource allocation.

A directed edge from vertex i to vertex j is weighted by the cross-channel gain from the transmitter of link i to the receiver of link j. This weight quantifies the interference caused by link i's transmission on link j's reception. The edge weight is a critical, physically meaningful parameter, often derived from channel estimation, that directly informs the message-passing process of the GNN.

The graph's topology is encoded in an adjacency matrix, where each entry Aij is the interference channel gain. Construction involves:

• Thresholding: Edges are only created if the interference exceeds a minimum power level, preventing a fully connected graph and reducing computational complexity.
• Path Loss Modeling: Gains are calculated using path loss and shadowing models based on the physical distance between transmitters and unintended receivers.

The interference graph is not static; it must be reconstructed periodically to reflect changes in the wireless environment. Triggers for reconfiguration include:

• User Mobility: New links are added or removed as devices move.
• Fading Events: Significant changes in channel gains due to large-scale fading.
• Scheduling Epochs: The graph is rebuilt at the start of each new scheduling interval to capture the latest network state for the GNN's decision-making.

Raw channel gains are often insufficient. Node and edge features are augmented with derived metrics to improve learning:

• Node Features: Signal-to-Noise Ratio (SNR), buffer status, traffic priority.
• Edge Features: Interference-to-Noise Ratio (INR), the product of cross-channel gain and the interfering link's transmit power.
• Global Features: A graph-level context vector can encode the total network spectral efficiency target or a fairness constraint.

In dense networks, constructing a fully connected interference graph is computationally prohibitive. A common scalability technique is to connect each vertex only to its k-nearest interfering neighbors. This sparsification is based on the assumption that far-away transmitters contribute negligible interference, preserving the most critical interference relationships while enabling the GNN to scale to large deployments.

A GNN architecture that models the wireless environment as a graph where nodes represent transmitters or frequency bins and edges represent interference or correlation relationships. The model learns to propagate information across the graph structure to solve tasks like spectrum mapping, resource allocation, and interference coordination. Unlike traditional grid-based methods, Spectrum GNNs naturally capture the irregular topology of wireless deployments and scale efficiently with network density.

A GNN-based approach where nodes in a spectrum graph iteratively exchange information with their neighbors to learn a global representation of the spectral environment. Each message passing round aggregates interference metrics from adjacent nodes, allowing the model to reason about cascading interference effects. This technique is foundational for distributed interference coordination, where each transmitter-receiver pair updates its understanding based on local neighborhood interactions without requiring centralized computation.

A discrete, learnable token representing an individual multipath component characterized by its delay, Doppler shift, and complex gain. When constructing an interference graph, each propagation path can be tokenized and fed into a transformer or GNN to process the channel as a set of paths rather than a monolithic matrix. This tokenization enables fine-grained modeling of how specific reflectors and scatterers contribute to mutual interference between links.

A learned vector representation that encodes the delay and Doppler shift characteristics of a propagation path. These embeddings serve as input tokens for transformer or GNN models processing channel responses in the delay-Doppler domain. In interference graph construction, Delay-Doppler embeddings capture the time-varying nature of interference, allowing the model to reason about how mobility and multipath fading affect link quality over time.

A mechanism that uses cross-attention to fuse information from two distinct signal representations, such as combining time-domain and frequency-domain features or merging outputs from multiple sensor modalities. In interference graph construction, cross-attention can fuse per-link channel measurements with global spectrum occupancy data, creating richer node features that capture both local link quality and the broader spectral context for more accurate interference weight estimation.

An attention mechanism that simultaneously models dependencies across spatial dimensions (antenna elements, node positions) and temporal dimensions (symbol periods, scheduling slots). When applied to interference graphs, this dual attention captures how interference patterns evolve over time due to mobility, fading, and dynamic scheduling decisions. The joint modeling enables GNN-based schedulers to anticipate future interference states rather than reacting only to current measurements.