Aristotle University Of Thessaloniki (AUTH)
Μανθάνοντα συστήματα ταξινομητών για εποπτευόμενα προβλήματα κατηγοριοποίησης
Abstract
dc:descriptionLearning Classifier Systems (LCS) belong to a class of Genetics-Based Machine Learning (GBML) systems, designed to work for both sequential and single-step problems. Although modern research in the LCS field has mainly focused on the case of sequential decision problems, during the 15 years, there has been a shift of interest toward applying LCS to single-step decision tasks, such as predictive Data Mining (DM), which has come to be considered one of the most important application domains for this class of systems. At the same time, while developments in the LCS research field have led to the maturity of associated technologies and promising implementations of new systems, it is clear that DM applications have not yet fully exploited the potential of LCS. The present thesis aims at addressing the corresponding challenges through a fruitful merging of both research areas, LCS and DM, especially focusing on the study of strength-based fitness in LCS. In this direction, the thesis presents a generalized supervised learning framework based on LCS (*S-LCS), independent of the employed fitness calculation schema, that is later equipped with a set of extensions to the traditional strength-based LCS framework, thus defining the SS-LCS algorithm. The behavior and performance of SS-LCS is studied in depth and compared to this of accuracy-based approaches, especially focusing on problems with extreme class imbalances. This study contributes to a better understanding of the learning dynamics involved in single-step classification problems and reveals the potential of supervised LCS in general, and strength-based ones more specifically, as real-world DM tools inducing tractable rule-based classification models, even in the presence of highly skewed class distributions. The present thesis also specifies and implements a general-use initialization component for the class of supervised LCS algorithms that extracts information about the structure of studied problems through a pre-training clustering phase and exploits this information by transforming it into rules suitable for the initialization of the learning process. The potential and robustness of our proposed approach are experimentally evaluated, revealing that it can lead to a significant improvement of the studied supervised LCS’ predictive and expressive abilities, as well as the required times for their training. In the last part of the thesis, we integrate the findings of all previous steps and, with the ultimate goal of revealing LCS potential for effective classification in real-world applications domains, we address the problems of: (a) predicting sparse episodes in environmental datasets and (b) multi-label classification. In both cases, it is experimentally proven that the proposed LCS approach presents clear advantages over conventional approaches, in terms of the predictive accuracy and understandability of the evolved knowledge representations.
Degree
thesis:*- Grantor dc:publisher
- Aristotle University Of Thessaloniki (AUTH)
- Year dc:date
- 2011
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Tzima, Fani
- Τζήμα, Φανή
Subjects
dc:subject × 20- Εξελικτική υπολογιστική
- Μανθάνοντα συστήματα ταξινομητών
- Κατηγοριοποίηση
- Αρχικοποίηση
- Αξιολόγηση ταξινομητών
- Πολυκατηγορική ταξινόμηση
- Ποιότητα αέρα
- Ανισόρροπες κατανομές κλάσεων
- Evolutionary computation
- Learning classifier systems (LCS)
- Classification
- LCS initialization
- Classifier evaluation
- Multi-label classification
- Air quality
- Unbalanced class distributions
- Engineering and Technology
- Electrical Engineering, Electronic Engineering, Information Engineering
- Επιστήμες Μηχανικού και Τεχνολογία
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/27952