Graph ML Foundations
Graph machine learning represents entities as nodes and relationships as edges. GNNs learn by propagating, weighting and transforming information over this connectivity. The topic is split into focused lessons so definitions, implementation decisions, diagnostics and common failure modes can be learned separately.
Learn the mechanism one decision at a time
Work through the lessons in order if the topic is new. If you already know the basics, open the specific leaf lesson that matches the operation, diagnostic or failure mode you need.
1Definition→
2Mechanism→
3Example→
4Diagnostic→
5Decision
Graph representationA graph contains nodes, edges and optional node/edge/global features. Direction, edge type and temporal structure may all matter. The important practical question is not only how the technique is defined, but what assumptions it introduces, which data are allowed to influence it, and how its effect should be validated on unseen evidence.
02Message passingA node receives messages from neighbours, aggregates them with a permutation-invariant operator, then updates its hidden representation. The important practical question is not only how the technique is defined, but what assumptions it introduces, which data are allowed to influence it, and how its effect should be validated on unseen evidence.
03Receptive fieldsAfter one layer, a node sees one-hop neighbours; deeper layers expand reach but can cause over-smoothing or over-squashing. The important practical question is not only how the technique is defined, but what assumptions it introduces, which data are allowed to influence it, and how its effect should be validated on unseen evidence.
04Architecture familiesGCN normalises aggregation; GraphSAGE samples/aggregates; GAT learns attention weights; GIN uses expressive sum aggregation; graph transformers add broader attention. The important practical question is not only how the technique is defined, but what assumptions it introduces, which data are allowed to influence it, and how its effect should be validated on unseen evidence.
05Graph tasksNode classification/regression, edge/link prediction and graph-level prediction require different splitting and readout strategies. The important practical question is not only how the technique is defined, but what assumptions it introduces, which data are allowed to influence it, and how its effect should be validated on unseen evidence.