Local structured learning for feature selection and causal discovery
Loading...
Authors
Gao, Tian
Issue Date
2016-05
Type
Electronic thesis
Thesis
Thesis
Language
ENG
Keywords
Electrical engineering
Alternative Title
Abstract
In this thesis, we introduce new methods to improve both the efficiency and accuracy of the existing local structure learning methods. Specifically, to improve the time efficiency of state-of-the-art methods, we first propose a constraint-based Simultaneous Markov Blanket Discovery (STMB) algorithm that exploits a newly discovered co-existence property among the local variables to remove the widely-used symmetry constraint. To improve accuracy, we introduce the Hybrid Markov Blanket(HMB) algorithm that combines STMB with a score-based search method to more accurately find the parent-and-child set within the local structure. Lastly, we develop a purely score-based algorithm, score-based Simultaneous Markov Blanket Discovery (S2TMB), and its more efficient variants to simultaneously improve the accuracy and efficiency of score-based MB discovery algorithms. We theoretically prove these algorithms’ soundness and completeness, and analyze their improved complexities. Experiments on both synthetic and standard datasets show the superior performance of these methods to the existing local structure discovery methods.
Description
May 2016
School of Engineering
School of Engineering
Full Citation
Publisher
Rensselaer Polytechnic Institute, Troy, NY