- Abstract: We present a scalable approach for unsupervised learning on graph-structured data based on a simple graph embedding learned via the triplet loss. For the community detection problem on the stochastic block model, our algorithm is optimal, with the same performance as the spectral technique based on the Bethe Hessian. On synthetic low-dimensional datasets, our algorithm generalizes well, having state of the art performances. In a semi-supervised learning framework, our algorithm extends naturally and incorporates the additional information with a great increase in performances.
- Keywords: clustering, community detection, graph embedding, stochastic block model
- TL;DR: A scalable algorithm for graph embedding achieving shown to be optimal on the stochastic block model.