Spatiotemporal GNN

The defining characteristic of a spatiotemporal GNN is its coupled architecture that processes two dimensions simultaneously:

• Spatial Dimension: Captured via graph convolutions (e.g., GCN, GAT) that aggregate features from a node's immediate neighbors, modeling interference or handover relationships.
• Temporal Dimension: Captured via recurrent units (e.g., GRU, LSTM) or temporal attention layers that learn patterns in historical sequences. This joint processing allows the model to understand not just where congestion is, but how it propagates through the network over time.

The mechanism for capturing time-series dynamics defines the model's forecasting horizon and computational cost:

• RNN-Based (LSTM/GRU): A GNN encodes each time step's graph into a node embedding, which is then fed sequentially into a recurrent neural network. Excellent for short-to-medium sequences but can suffer from vanishing gradients.
• Temporal Convolutional Networks (TCN): Apply 1D dilated convolutions along the time axis for each node, enabling parallel computation over long sequences with a fixed receptive field.
• Temporal Attention (Transformer): Uses self-attention across time steps, allowing the model to directly weigh the importance of any historical point. This is the dominant approach for capturing long-range dependencies.

Spatiotemporal models must handle the evolution of the graph itself, not just node features:

• Static Graph, Dynamic Features: The most common assumption in cellular traffic forecasting. The base station topology is fixed, but traffic load and user counts change over time.
• Discrete-Time Dynamic Graph: A sequence of graph snapshots where edges appear or disappear at each time step, modeling user handovers or intermittent interference.
• Continuous-Time Dynamic Graph: Models events (edge additions/deletions) as they occur on a continuous timeline using neural ordinary differential equations or temporal point processes, ideal for high-resolution mobility modeling.

The canonical application is predicting future traffic load across a network of base stations:

• Input: A sequence of historical traffic snapshots, each represented as a graph where nodes are base stations and edges represent spatial proximity or handover adjacency.
• Output: A multi-step prediction of traffic volume for every base station in the grid.
• Key Benefit: By learning spatial diffusion patterns, the model accurately forecasts cascading congestion—where a traffic spike in one cell predictably migrates to adjacent cells over subsequent time intervals.

Modern architectures decouple spatial and temporal attention for computational efficiency:

• Temporal Self-Attention: For a single node, computes attention scores across all historical time steps to capture periodicity and trends.
• Spatial Self-Attention: For a single time step, computes attention scores across all nodes in the graph to capture instantaneous spatial dependencies.
• Interleaved vs. Parallel: Some models apply spatial then temporal attention sequentially, while others compute them in parallel and fuse the representations. The parallel approach is faster but may miss subtle cross-dimensional interactions.

In a cellular network, the properties of a connection between two base stations are not static. A spatiotemporal GNN must model how edge features change over time:

• Dynamic Path Loss: Signal attenuation between two cells changes due to environmental factors like moving vehicles or weather.
• Time-Varying Interference: The interference relationship between two cells depends on their current transmission schedules and user distributions.
• Edge RNNs: A dedicated recurrent unit can be assigned to each edge to track its state evolution independently, feeding this information into the spatial message-passing step.

A deep learning architecture designed to operate directly on graph-structured data, learning representations of nodes, edges, or entire graphs by recursively aggregating information from local neighborhoods. In cellular networks, a GNN processes the interference graph where nodes represent base stations and edges represent significant radio relationships. The core operation involves each node receiving messages from its neighbors, transforming them with a learnable function, and updating its own state—a process that naturally respects the permutation invariance of the network topology.

A GNN variant designed to process graphs whose topology or node features evolve over time, making it essential for modeling user mobility and changing traffic patterns in cellular networks. Unlike static GNNs that assume a fixed graph structure, dynamic GNNs handle edge additions and deletions—such as users moving between cells or new interference relationships forming. Architectures include temporal graph networks that maintain memory states for nodes and evolving graph convolutions that apply recurrent updates to both structure and features.

A GNN architecture that introduces a self-attention mechanism to dynamically weigh the importance of different neighboring nodes during feature aggregation. Rather than treating all neighbors equally, GAT computes attention coefficients that allow the model to focus on the most relevant connections—critical in cellular topologies where some interfering base stations matter far more than others. Multi-head attention further stabilizes learning by running multiple attention mechanisms in parallel and concatenating their outputs.

A general framework that formalizes how information propagates across graph topology. Each node iteratively updates its state by receiving and aggregating messages from neighboring nodes through three functions: a message function that computes what information to send, an aggregation function (sum, mean, or max) that combines incoming messages, and an update function that produces the new node state. MPNNs unify many GNN variants under a common mathematical framework and are the theoretical backbone of spatiotemporal models.

A GNN architecture that applies the global self-attention mechanism of a Transformer to all nodes in a graph, enabling each node to attend to every other node regardless of graph distance. This directly addresses the over-squashing problem where information from distant nodes is compressed into fixed-size vectors. Graph Transformers require positional encoding to inject topological information since self-attention is structure-agnostic. They excel at capturing long-range dependencies in large cellular deployments.

A graph representation of a wireless network where nodes represent base stations or user equipment and edges represent significant radio relationships such as interference, handover adjacency, or signal dominance. The topology is typically constructed from drive-test data, propagation models, or real-time measurements. Edge features encode path loss and channel gain, while node features capture transmission power, load, and queue length. This graph serves as the spatial backbone for spatiotemporal GNNs.