Learning Markov networks with context- specific independences

EDERA, ALEJANDRO; SCHLÜTER, FEDERICO; BROMBERG, FACUNDO

INTERNATIONAL JOURNAL ON ARTIFICIAL INTELLIGENCE TOOLS

WORLD SCIENTIFIC PUBL CO PTE LTD

Lugar: London, UK; Año: 2014 vol. 23

0218-2130

This work focuses on learning the structure of Markov networks from data. Markov networks are parametric models for compactly representing complex probability distributions. These models are composed by: a structure and numerical weights, where the structure describes independences that hold in the distribution. Depending on which is the goal of structure learning, learning algorithms can be divided into: density estimation algorithms, where structure is learned for answering inference queries; and knowledge discovery algorithms, where structure is learned for describing independences qualitatively. The latter algorithms present an important limitation for describing independences because they use a single graph; a coarse grain structure representation which cannot represent flexible independences. For instance, context-specific independences cannot be described by a single graph. To overcome this limitation, this work proposes a new alternative representation named canonical model as well as the CSPC algorithm; a novel knowledge discovery algorithm for learning canonical models by using context-specific independences as constraints. On an extensive empirical evaluation, CSPC learns more accurate structures than state-of-the-art density estimation and knowledge discovery algorithms. Moreover, for answering inference queries, our approach obtains competitive results against density estimation algorithms, significantly outperforming knowledge discovery algorithms. Read More: http://www.worldscientific.com/doi/abs/10.1142/S0218213014600306