Please use this identifier to cite or link to this item:
Title: Probabilistic learning: Sparsity and non-decomposable losses
Authors: YE NAN
Keywords: statistical learning, sparsity, non-decomposable losses, conditional random fields, F-measures
Issue Date: 11-Jan-2013
Citation: YE NAN (2013-01-11). Probabilistic learning: Sparsity and non-decomposable losses. ScholarBank@NUS Repository.
Abstract: This thesis considers dealing with increasingly more complex data and performance measures in statistical learning. For structured data, we consider conditional random fields (CRFs) with sparse potential functions. For a class of sparse high-order CRFs and a class of sparse factorial CRFs, we give polynomial time exact inference and learning algorithms, and show that they perform well on synthetic and real datasets. For general loss functions, we focus learning to optimize the F-measures with the Empirical utility maximization (EUM) approach and the decision-theoretic approach (DTA) approach. Theoretically, EUM is consistent, and given accurate models, they are asymptotically equivalent on large training and test sets. Empirically, EUM appears to be more robust against model misspecification, whereas given a good model, DTA appears to be better for rare classes. We give a $O(n^2)$-time algorithm to compute predictions with optimal expected F-measures for independent labels.
Appears in Collections:Ph.D Theses (Open)

Show full item record
Files in This Item:
File Description SizeFormatAccess SettingsVersion 
thesis.pdf1 MBAdobe PDF



Page view(s)

checked on Dec 16, 2018


checked on Dec 16, 2018

Google ScholarTM


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.