Concept Drift Detection in Dynamic Probabilistic Relational Models

Dynamic probabilistic relational models, which are factorized w.r.t. a full joint distribution, are used to cater for uncertainty and for relational and temporal aspects in real-world data. While these models assume the underlying temporal process to be stationary, real-world data often exhibits non-stationary behavior where the full joint distribution changes over time. We propose an approach to account for non-stationary processes w.r.t. to changing probability distributions over time, an effect known as concept drift. We use factorization and compact encoding of relations to efficiently detect drifts towards new probability distributions based on evidence.

Dynamic Domain Sizes in Temporal Probabilistic Relational Models

Probabilistic dynamic relational models (PDRMs) allow for an expressive, yet sparse and efficient representation of uncertain temporal (dynamic) and relational information with a fixed (static) set of domain objects (entities). While for different points in time, information about objects may differ, the set of objects under consideration is the same for all time points in standard PDRMs. Motivated by examples from a logistics application, in this paper we extend the theory of PDRMs with dynamically changing sets of domain objects. The paper introduces the semantics of so-called PD2RMs and analyses model management as well as query answering problems and algorithms.

DARPA’s Role in Machine Learning

Machine learning methods provide a way for artificial intelligence systems to learn from experience. This article describes four threads of machine learning research supported and guided by the Defense Advanced Research Projects Agency — probabilistic modeling for speech recognition, probabilistic relational models, the integration of multiple machine learning approaches into a task-specific system, and neural network technology. These threads illustrate the Defense Advanced Research Projects Agency way of creating timely advances in a field.