Graph Neural Network-based framework for predicting top Betweenness Centrality Nodes across Unseen Graphs

Betweenness centrality is a key indicator for quantifying node importance in complex networks, but remains computationally expensive for large-scale graphs due to its global dependency on all-pairs shortest paths. This research describes a machine learning-based methodology for estimating betweenness centrality with low-complexity topological features and graph neural networks. We use the GraphSAGE algorithm with two distinct neighborhood aggregation schemes- mean and max aggregator to create predictive models that generalize across unseen nodes. Four synthetic network models- Erdős- Rényi, Watts-Strogatz, Barabási-Albert, and Newman-Park are generated to evaluate generalization across diverse topologies. The experiments involve training GraphSAGE models on smaller graphs and testing on larger unseen graphs from the same topology. The Max aggregator achieves an average of approximately 3% higher Top-10 node overlap and 1% higher Spearman correlation compared to the Mean aggregator across all datasets resulting in high rank correlations and identification of most important nodes.