GraphSAGE

Before GraphSAGE, graph neural networks were mostly transductive: they memorized embeddings for a fixed graph and could not handle new nodes. GraphSAGE changed the framing from learning node embeddings to learning an aggregation function. For each node it samples a fixed-size neighborhood, aggregates neighbor features with a learnable function (mean, LSTM, or pooling), and concatenates the result with the node's own features. Because the model only learns how to aggregate, it works inductively on entirely unseen graphs and scales to billions of edges with minibatch training.\n\nThe NeurIPS 2017 paper reports supervised micro-F1 of 0.839 on citation graphs, 0.954 on the Reddit post graph and 0.612 on the PPI protein graph, roughly 40-60% relative gains over raw-feature and DeepWalk baselines. The design was validated at industrial scale by PinSAGE, Pinterest's production recommender built directly on these ideas. GraphSAGE's neighborhood-sampling recipe became the template for nearly all scalable GNN training that followed, including GraphSAINT and Cluster-GCN.

Paper-reported (supervised micro-F1): 0.839 citation graphs, 0.954 Reddit, 0.612 PPI (Hamilton et al., NeurIPS 2017); ~40-60% relative gains over raw-feature and DeepWalk baselines. Industrial validation via PinSAGE at Pinterest.