Fraud rings, recommendations, knowledge graphs, hosts-and-flows in network security — most interesting data is a graph, and the tell is often in the structure, not the rows. One repeated move powers it all: message passing — every node aggregates its neighbours and updates itself with a shared rule. Learn it visibly, then scale it (GraphSAGE), sharpen it (GAT), and prove on the capstone that structure beats a features-only baseline.
Finish the GNN fraud capstone (GCN from scratch), then redo it in PyTorch Geometric with GraphSAGE/GAT and compare against a features-only baseline — proving structure adds signal. For the cyber track: model hosts/flows as a graph and detect lateral-movement clusters. Read the GCN or GraphSAGE paper with your research-skills method. You've got GNNs when you can explain message passing, implement a GCN, train node classification in PyG, and articulate when graph structure beats tabular features.
Guides: gnn-fraud capstone ↗ · module-c guide ↗