Structure Learning for Statistical Relational Models

Many data sets routinely captured by businesses and organizations are relational in nature yet over the past decade most machine learning research has focused on “flattened” propositional data. Propositional data record the characteristics of a set of homogeneous and statistically independent objects; relational data record characteristics of heterogeneous objects and the relations among those objects. Examples of relational data include citation graphs, the World Wide Web, genomic structures, fraud detection data, epidemiology data, and data on interrelated people, places, and events extracted from text documents. Relational data offer unique opportunities to boost the accuracy of learned models and improve the quality of decision-making if the algorithms can learn effectively from the additional information the relationships provide.