Discriminative Probabilistic Models for Relational Data

摘要: In many supervised learning tasks, the entities to be labeled are related each other in complex ways and their labels not independent. For example, hypertext classification, of linked pages highly correlated. A standard approach is classify entity independently, ignoring correlations between them. Recently, Probabilistic Relational Models, a relational version Bayesian networks, were used define joint probabilistic model for collection entities. this paper, we present an alternative framework that builds on (conditional) Markov networks addresses two limitations previous approach. First, undirected models do impose acyclicity constraint hinders representation important dependencies directed models. Second, well suited discriminative training, where optimize conditional likelihood given features, which generally improves classification accuracy. We show how train these effectively, use approximate inference over learned collective multiple provide experimental results webpage task, showing accuracy can significantly improved by modeling dependencies.