University of Konstanz
Graduiertenkolleg / PhD Program
Computer and Information Science

Colloquium of the Department and the PhD Program

title

Infinite hidden relational models

speaker

Dr. Volker Tresp, Siemens, Munich, Germany
München, Germany

date & place

Wednesday, 14.02.2007, 16:15 h
Room C252

abstract

Relational learning is an object oriented approach to representation and learning that clearly distinguishes between entities (e.g., objects), relationships and their respective attributes and represents an area of growing interest in machine learning. In my presentation I will present the infinite hidden relational model (IHRM), which describes a relational model as a random field quite similar to the random (hidden) Markov fields known from vision and other applications. In the IHRM, a latent variable is introduced for each object or entity. The state of a latent variable reflects, first, the local attributes of the entity, second, the properties of the relations the entity is directly involved in and, third, the latent states of the latent variables of the other entities involved in the relations. In the network of latent variables, information can thus propagate reducing the need for extensive structural model selection, typical for other approaches to relational learning. In relational modeling, different classes of entities are involved requiring a class-specific complexity in the latent representation. Therefore it is sensible to work with Dirichlet process (DP) mixture models in which each entity class can optimize its own representational complexity in a self-organized way. The resulting representation is a network of interacting DPs.

The IHRM presents a generic approach to relational modeling and should thus be widely applicable. In the presentation I will provide experimental results for a recommendation system and for the prediction of gene functions using relational information.