Node Feature Engineering

A feature must directly correlate with the optimization objective. For a base station node, transmission power and current load are highly discriminative for energy-saving tasks, while a static site ID is not.

• Good features: Queue length, PRB utilization ratio, average channel quality indicator (CQI)
• Poor features: Random noise, arbitrary node IDs, constant values
• The feature set must allow the GNN to distinguish between a congested macro-cell and an idle small-cell.

GNNs are sensitive to the scale of input data. Features with vastly different ranges (e.g., power in dBm vs. load as a percentage) can destabilize gradient descent.

• Apply Z-score normalization or min-max scaling to ensure zero mean and unit variance.
• Handle outliers: A single faulty sensor reporting extreme power can skew the entire feature distribution.
• Use logarithmic scaling for wide-range metrics like interference power to compress dynamic range.

The ordering of nodes in the input matrix is arbitrary. The feature vector for a base station must not depend on its index in a list. The GNN architecture handles permutation invariance, but the features themselves must be intrinsic properties of the node.

• A feature like 'distance to sector center' is valid.
• A feature like 'index in the adjacency matrix' is invalid and breaks the model's ability to generalize to new network topologies.

Cellular networks are dynamic. A static snapshot of load is less informative than a time-series window. For spatiotemporal GNNs, features should encode recent history.

• Use exponentially weighted moving averages of load over the last T seconds.
• Include first-order derivatives (rate of change) to capture traffic surges.
• A feature vector might be: [current_load, avg_load_5min, load_trend, peak_load_1hr].

Raw features often fail to capture a node's structural role. Augmenting features with graph-theoretic measures provides crucial context for message passing.

• Node degree: Number of significant interferers or handover neighbors.
• Betweenness centrality: How often a node lies on the shortest path between others, indicating a routing or backhaul bottleneck.
• Clustering coefficient: The density of interference in a node's local neighborhood, useful for identifying dense urban canyons.

Real-world network telemetry is often incomplete. A feature engineering pipeline must be robust to missing sensor readings without introducing statistical bias.

• Imputation strategy: Replace missing values with the global mean or a learned embedding vector, not zero (which is a valid power level).
• Use a binary mask feature to explicitly signal to the GNN that a value was imputed.
• Avoid features with >50% missing rate across the topology, as they introduce noise rather than signal.