English Español Français Deutsch
← Back to the list
Foundational figures Foundational figures Methodologically adjacent

Jure Leskovec

What graph structure adds to the kinds of reasoning and retrieval problems language models currently handle without it — and what gets missed when relationships in data are flattened into text.

Leskovec's node2vec (2016) was one of the first methods to make graph-structured data work as input to standard ML pipelines — embed nodes as vectors that preserve neighborhood information, then use the embeddings as features anywhere a vector goes. The Stanford Network Analysis Project (SNAP) datasets and methods that followed became the default substrate for graph-ML research. For ai100, where the question of "how does a model know what brand to mention" sits in graph-structured-reasoning territory — which entities are connected to which contexts in training data — the graph perspective is a usefully orthogonal angle on what current LLM evaluation tends to measure only on the surface.

Worth following when
you want to think about LLM knowledge and reasoning behavior in terms of graph structure underneath the text, rather than as a property of the text itself.
Topics
graph machine learning and node embeddings (node2vec); the Stanford Network Analysis Project (SNAP); graph-structured reasoning as a lens on LLM behavior.
Key works
node2vec: Scalable Feature Learning for Networks (2016, co-author); Stanford Network Analysis Project datasets and tools (2007 onward); graph-ML textbook and curriculum contributions.