Knowledge discovery in networks

Networks are an increasing common method of representing the relationships among sets of interacting entities. This basic data structure is reflected in how we analyze and understand social networks, networks of scholarly citations and web pages, and networks of computers and communications devices. Over the past five years, my students and I have developed a number of methods for learning statistical models of networks that can make accurate predictions about the attributes of nodes in the network and also provide insight into the broad structure of statistical dependencies among different types of nodes. These models build on methods developed previously in statistics, machine learning, and knowledge discovery, including Bayesian networks and probability estimation trees.

We have applied these techniques to a wide variety of problems, including citation analysis and fraud detection. Most recently, we have applied these techniques to detect fraud among stock brokers, in a joint project with the National Association of Securities Dealers. We have also developed an open-source software environment incorporating our tools for statistical modeling and ad hoc querying of relational data.