Learning Statistical Patterns in Relational Data Using Probabilistic Relational Models

摘要

This report describes techniques for learning probabilistic models of relational data, and using these models to interpret new relational data. This effort focused on developing undirected probabilistic models for representing and learning graph patterns, learning patterns involving links between objects, learning discriminative models for classification in relational data, developing and labeling two real-world relational data sets - one involving web data and the other a social network - and evaluating the performance of our methods on these data sets, and dealing with distributions that are non-uniform, in that different contexts (time periods, organizations) have statistically different properties. The technology developed under this effort was transitioned and is being used under the Perceptive Assistant Program (PAL) at DARPA.