Back to search

Aristotle University Of Thessaloniki (AUTH)

Μηχανές διανυσμάτων υποστήριξης για θορυβώδεις και ασαφείς ετικέτες

Abstract

dc:description

This dissertation studies the problem of pattern recognition with noisy labels or fuzzy training instances. Label noise refers to wrong label assignments whereas fuzziness refers to instances with different membership values for their class. Such problems can be solved effectively by modelling the noise or the fuzziness of each training instance and using instance-weighted learning algorithms. We identify two major challenges: (a) how to model instance label noise or fuzziness; and (b) how to produce optimal instance weights using these models. To overcome these challenges, this work proposes strategies based on variants of the Support Vector Machine (SVM) algorithm that feature instance weights. The main theoretical contribution of this work is a set of new hyper-parameter selection methods for noisy and fuzzy training data. First, we extend the span error bound theory of Vapnik and Chapelle to the fuzzy SVM algorithm and we prove an unbiased estimator and an upper bound of the test error. Also, we define estimators of the test error in problems where the probability of correct label assignments can be estimated and we prove their unbiasedness. The work proposes and evaluates strategies for training image classifiers using label-noisy clickthrough data. The strategies include automatic training set construction, label noise modelling and instance weights generation for weighted SVM variants. Also, the new hyper-parameter selection methods are used to perform instance weights selection for fuzzy SVM classifiers. Furthermore, the work proposes and evaluates strategies for robust online classifier training using streaming clickthrough data. The online classifiers are designed to be theoretical equivalent to the batch-trained ones that use explicit instance weights. Finally, the work proposes a strategy for training data-driven models for human behaviour interpretation in domains with little or none medical guidelines. Based on this, the work describes methods for identifying individuals at risk from measurements of eating behaviour during meals. Domain experts assign membership values and the optimal weights are selected using the new methods. The experimental evaluation in the above applications demonstrates that modelling noise and fuzziness in combination with instance weights produced using the new hyper-parameter selection methods leads to better generalisation performance and robustness than non-weighted solutions.

Degree

thesis:*
Grantor dc:publisher
Aristotle University Of Thessaloniki (AUTH)
Year dc:date
2018

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Sarafis, Ioannis
  • Σαράφης, Ιωάννης

Subjects

dc:subject × 14

Rights

dc:rights
Statement dc:rights
  • BY_SA
Language dc:language
gre

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/44863

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Sarafis, Ioannis; Σαράφης, Ιωάννης. Μηχανές διανυσμάτων υποστήριξης για θορυβώδεις και ασαφείς ετικέτες. Aristotle University Of Thessaloniki (AUTH), 2018. http://hdl.handle.net/10442/hedi/44863