Learning thin junction trees via graph cuts

Dafna Shahaf*, Anton Chechetka, Carlos Guestrin

*Corresponding author for this work

Research output: Contribution to journalConference articlepeer-review

13 Scopus citations

Abstract

Structure learning algorithms usually focus on the compactness of the learned model. However, for general compact models, both exact and approximate inference are still NP-hard. Therefore, the focus only on compactness leads to learning models that require approximate inference techniques, thus reducing their prediction quality. In this paper, we propose a method for learning an attractive class of models: bounded-treewidth junction trees, which permit both compact representation of probability distributions and efficient exact inference. Using Bethe approximation of the likelihood, we transform the problem of finding a good junction tree separator into a minimum cut problem on a weighted graph. Using the graph cut intuition, we present an efficient algorithm with theoretical guarantees for finding good separators, which we recursively apply to obtain a thin junction tree. Our extensive empirical evaluation demonstrates the benefit of applying exact inference using our models to answer queries. We also extend our technique to learning low tree-width conditional random fields, and demonstrate significant improvements over state of the art block-L1 regularization techniques.

Original languageAmerican English
Pages (from-to)113-120
Number of pages8
JournalProceedings of Machine Learning Research
Volume5
StatePublished - 2009
Externally publishedYes
Event12th International Conference on Artificial Intelligence and Statistics, AISTATS 2009 - Clearwater, FL, United States
Duration: 16 Apr 200918 Apr 2009

Fingerprint

Dive into the research topics of 'Learning thin junction trees via graph cuts'. Together they form a unique fingerprint.

Cite this