Paper Summary - Probabilistic relational models

D. Koller. Probabilistic relational models. In S. Dzeroski and P. Flach, editors, Proceedings of Ninth International Workschop on Inductive Logic Programming (ILP-
1999). Springer, 1999.


This paper highlights deficiencies that exist with Bayesian networks; mainly that BNs cannot represent complex domains because they cannot represent models of domains that they are not aware of in advance. They present probabilistic relational models as a language for describing probabilistic models. Entities, their properties and relations are represented with the language.

The key points in this paper are:

• BNs are very useful and have been successful as a way to perform probabilistic reasoning
  
• BNs are inadequate for representing large complex domains


• BNs lack the concept of an object and therefore there is no concept of similarity among objects across contexts
  


• Probabilistic relational models extend BNs by adding concepts of individuals, properties and relations

Objects are the 'basic entities' in a PRM and partitioned into disjoint classes with a set of associated attributes. In addition to objects, relations also make up the vocabulary. It states that the goal of the PRM is to define a probability distribution over a set of instances for a schema.

The key distinctions here with PRMs and Bayesian networks are PRMs define the dependency model at the class level and they use the relational structure of the model. They are more expressive then Bayesian networks.

Regarding inference, they show how this expressiveness helps rather than further complicates the processing.

The paper is dense but interesting.