Section: New Results

Probabilistic models for large graph

We have developped new approaches for the statistical analysis of large-scale undirected graphs. The main insight is to exploit the spectral decomposition of subgraph samples, and in particular their Fiedler eigenvalues, as basic features for density estimation and probabilistic inference. Our contributions are twofold. First, we develop a conditional random graph model for learning to predict links in information networks (such as scientific coauthorship and email communication). Second, we propose to apply the resulting model to graph generation and link prediction. This work is to published in the Journal of Machine Learning Research, the top journal in the field of machine learning.