Spatial Graph Convolution

The core operation computes a node's new representation by aggregating the feature vectors of its immediate neighbors and combining them with its own features. This is a localized, permutation-invariant operation.

• Mean Aggregation: Takes the element-wise average of neighbor features.
• Sum Aggregation: Sums neighbor features, preserving neighborhood size information.
• Max Aggregation: Takes the element-wise maximum, capturing the most salient signals.

The aggregated vector is typically passed through a learnable weight matrix and a non-linear activation function like ReLU to produce the updated node embedding.

Unlike spectral methods that require the entire graph Laplacian, spatial convolution operates strictly on a node's k-hop neighborhood. The receptive field is defined by the number of successive convolutional layers.

• 1 Layer: Aggregates from direct neighbors.
• 2 Layers: Aggregates from neighbors of neighbors.
• Scalability: Computation is linear with respect to the number of edges, making it suitable for massive, sparse cellular topologies. This locality mirrors the sliding window of a CNN, where a pixel's context is defined by its spatial proximity.

Spatial methods are inherently inductive, meaning they can generate embeddings for nodes never seen during training. Because the convolution is defined as a function of local neighborhood features, the learned weight matrix can be applied to any new node added to an existing graph or to an entirely new graph.

• Dynamic Networks: Essential for cellular networks where new user equipment constantly joins.
• Generalization: A model trained on one city's topology can be deployed in another without retraining. This contrasts with transductive spectral methods that require a fixed graph structure for the Laplacian eigenbasis.

To handle large graphs where full neighborhood aggregation is computationally prohibitive, spatial convolution is paired with neighborhood sampling. Instead of aggregating all neighbors, a fixed-size subset is randomly sampled for each node during training.

• GraphSAGE: A seminal framework that popularized sampling for spatial convolution.
• Mini-Batch Training: Enables training on graphs with billions of edges by controlling the memory footprint.
• Variance Control: Advanced samplers use importance weighting to reduce the variance introduced by the stochastic sampling process.

Spatial convolution naturally incorporates edge features into the message-passing process. In a cellular topology, an edge's path loss or interference level is a critical feature.

• Message Function: A neural network takes the source node feature and the edge feature to compute a message: m_ij = MLP(h_i, e_ij).
• Weighted Aggregation: Edge features can act as gating mechanisms or attention coefficients, allowing the model to discount messages from distant or weakly connected neighbors. This provides a direct mechanism for encoding radio propagation physics into the neural network.

A spatial convolution layer is mathematically permutation invariant by design. The aggregation function (sum, mean, max) is a symmetric operation, meaning the output for a node is independent of the arbitrary order in which its neighbors are processed.

• No Node Ordering: The graph does not require a canonical node ordering, unlike images.
• Stable Representations: The same network topology will always produce the same embedding regardless of how the adjacency list is stored in memory. This property is a fundamental requirement for any valid graph neural network operation.

The general computational framework that formalizes spatial graph convolution. An MPNN operates in two phases: a message function computes information to send from a neighbor v to a target node u, and an update function aggregates all incoming messages to produce a new state for u. This iterative process allows information to propagate across the graph's topology. In a cellular interference graph, a message might encode a neighboring base station's current transmit power and path loss, while the update function combines these to adjust the target cell's resource allocation strategy.

A spatial convolution variant that introduces a self-attention mechanism to dynamically weight the importance of neighboring nodes during aggregation. Instead of treating all neighbors equally, a GAT computes an attention coefficient for each edge, allowing the model to focus on the most critical interference relationships. For example, a base station node can learn to assign a higher weight to a dominant interferer and ignore a negligible one, leading to more nuanced and effective power control decisions in dense heterogeneous networks.

An inductive spatial convolution framework designed for scalability and generalization. GraphSAGE learns a set of aggregator functions (e.g., mean, LSTM, pooling) to combine features from a fixed-size sample of a node's local neighborhood, rather than the full neighborhood. This sampling approach controls computational cost on massive cellular topologies and, crucially, enables the model to generate embeddings for previously unseen base stations or user equipment added to the network without retraining the entire model.

A critical failure mode in deep spatial GNNs where node representations become indistinguishable after too many convolution layers. As the receptive field expands, each node's embedding converges to a global average, losing all local discriminative information. In a cellular network, this means a congested micro-cell and an idle macro-cell would appear identical to the model. Mitigation strategies include using skip connections (e.g., GraphSAGE), residual layers, or limiting the model depth to prevent excessive information diffusion.

A phenomenon where information from a large, exponentially-growing receptive field is compressed into a fixed-size feature vector, preventing the GNN from learning long-range dependencies. In a large cellular topology, a spatial convolution may fail to propagate a critical congestion signal from a distant edge cell to a central controller because the message is 'squashed' through too many narrow bottlenecks. Graph rewiring techniques or Transformer-based architectures are often employed to mitigate this bottleneck.

A fundamental training technique for scaling spatial convolution to planet-scale cellular graphs. Instead of performing a full recursive neighborhood expansion for each node in a mini-batch, sampling methods randomly select a subset of neighbors at each hop. This bounds the computational and memory footprint. Common strategies include uniform sampling, importance sampling based on node degree, and cluster-based sampling that partitions the graph into dense subgraphs to maintain local structure while drastically reducing batch computation time.