Learning High-Dimensional Mixtures of Graphical Models

Abstraction: Efficient unsupervised learning of mixtures of discrete graphical models via tree approximation

Key points:

• Problem: unsupervised estimation of mixtures where each component has a potentially different Markov graph structure and parameters
• Output is a tree-mixture model serving as a good approximation to the underlying graphical model mixture
• Method is efficient when the union graph (union of all component Markov graphs) has sparse vertex separators
• Covers tree mixtures and mixtures of bounded-degree graphs; proves correct recovery of union graph structure
• Sample and computational complexities scale as poly(p, r) for r-component mixture of p-variate graphical models
• Extended to mixtures of locally tree-like graphs under correlation decay regime

Connections: Arxiv · Graphical Models · Mixture Models · Structure Learning · Markov Random Fields

Source: http://arxiv.org/abs/1203.0697