In an earlier lecture, we saw how Graph Convolutional Networks propagate information through a normalized adjacency matrix. That formulation is elegant and mathematically grounded, but it comes with an important limitation: it is fundamentally tied to the specific graph on which it is trained.
- What if new nodes appear after training?
- What if we want to apply the model to a different graph with similar structure?
- What if the graph is too large to process all at once?
GraphSAGE was introduced to address precisely these challenges. Instead of relying on a fixed, globally normalized adjacency matrix, GraphSAGE learns how to aggregate information from a node’s neighbors in a way that generalizes to unseen nodes and even unseen graphs.
