Apparaît dans la collection : Structured Regularization Summer School - 19-22/06/2017

Many statistics and unsupervised learning problems can be formalized as estimating a structured matrix or a graph from noisy or incomplete observations. These problems present a large variety of challenges, and an intriguing interplay between computational and statistical barriers. I will provide an introduction to recent work in the area, with an emphasis on general methods and unifying themes. 1) Random matrix theory and spectral methods. 2) The semidefinite programming approach to graph clustering. 3) Local algorithms and graphical models. The hidden clique problem. 4) Non-negative matrix factorization.