Research in distributed networks earns Notable Paper Award at AISTATS

The research provides a way to efficiently reveal relationships between even distant entities in a network.

zhaoshi meng dr dennis wei and prof al hero Enlarge
Left to right: Zhaoshi Meng, Dr. Dennis Wei, and Prof. Al Hero

Al Hero, III, R. Jamison and Betty Williams Professor of Engineering, received a Notable Paper Award along with graduate student Zhaoshi Meng and postdoctoral researcher Dennis Wei, for the paper, “Distributed Learning of Gaussian Graphical Models via Marginal Likelihoods.” The paper will be presented at the Sixteenth International Conference on Artificial Intelligence and Statistics (AISTATS), held April 29-May 1, 2013 in Scottsdale, AZ. The research provides a way to efficiently reveal relationships between even distant entities in a network, whether it be a social network or a network of sensors.

Describing the research in general terms, Prof. Hero states that, “Many kinds of networks, whether networks of sensors or networks of people, are often described in terms of a graphical model representing statistical correlations between agents in the network. Characterizing the dependencies between any two agents usually requires data from the full network. In a social network for example, correlation between two people might be attributed to either direct interaction or indirect interaction through a chain of friends connecting them. Distinguishing between direct and indirect interaction requires data from the full network.”

However, determining that relationship in large, distributed networks is often difficult due to the prohibitive computation and communication costs involved. Instead, said Hero, “we propose a distributed approach to the estimation of correlations that decomposes the problem into small, local problems solved by each agent communicating only with its neighbors.”

The authors discovered that by looking at individuals or nodes in a distributed network that were no further apart than two degrees of separation, they could use that information to estimate the relationship between even disparate nodes to nearly the same level of accuracy as a much more expensive centralized estimator that looks at the entire network.

“We provide theoretical guarantees for this behavior and empirical illustrations in various forms of networks,” stated Prof. Hero.

The research was supported by the Army Research Office under the MURI, “Value-centered Information Theory for Adaptive Learning, Inference, Tracking, and Exploitation.”

AISTATS brings together researchers at the intersection of computer science, artificial intelligence, machine learning, statistics, and related areas. Similarly, Prof. Hero takes a highly interdisciplinary approach to his research, working with investigators across many disciplines. In addition to being a professor of Electrical Engineering and Computer Science, he is a professor of Statistics and of Biomedical Engineering.