A visual companion to the episode: how biased random walks turn a graph into pseudo-language, how p and q reshape context, and why node2vec became the practical shallow baseline before graph neural methods took over.
1607.00653
Theme: word2vec → random walks → node vectors
node2vec keeps the word2vec objective and rewires the context generator. The page below draws the full pipeline as a moving graph system rather than a text recipe.
This second-order transition depends on the previous node. Hover the heatmap and the graph to see how return, in-neighborhood, and outward moves reweight context under different p and q settings.
The paper’s influence came from useful benchmark improvements. The harder claim is stronger: whether those gains isolate flexible role-versus-community capture, or mostly show that biased sampling is a better shallow baseline.
This mock embedding map contrasts two geometric stories. In one, nodes of the same community bunch together. In the other, far-apart hubs or bridges can align by role even when their local neighborhoods differ.
Aditya Grover, Jure Leskovec, 2016 · arXiv 1607.00653
Bryan Perozzi, Rami Al-Rfou, Steven Skiena, 2014 · predecessor walk corpus idea
Jian Tang et al., 2015 · shallow embedding baseline focused on proximity
William Hamilton, Rex Ying, Jure Leskovec, 2017 · later inductive shift in graph learning
Yoav Goldberg, Omer Levy, 2014 · skip-gram with negative sampling intuition
Keith Henderson et al., 2012 · structural-role framing behind the paper’s stronger claim