Back to search

University of Patras

Νέοι αλγόριθμοι μηχανικής μάθησης για την επαγωγή γνώσης από εκπαιδευτικά / μαθησιακά δεδομένα

Abstract

dc:description

Nowadays, large amount of educational data is becoming more and more available. Most of these data emanate from institutional student information systems, virtual learning environments, attendance monitoring systems and library systems. These systems can record any student activity that is supported, such as reading, writing, exams taking, tasks performed and peer communications, information on staff, content and the institution and so forth. The analysis of these datasets attracted much attention among researchers for its potential to identify patterns, trends and predictions that can be used to optimize the learning process and its outcomes. In the field of Education, Educational Data Mining and Learning Analytics have expanded rapidly, emerging as dynamic interdisciplinary fields of study. Their ultimate common goal is to understand and improve the learning process and environment.Predicting students’ progression and learning outcomes is one of the most important task in the field of Educational Data Mining and Learning Analytics. Regardless of how the problem is being approached, its solution is of high practical value, as it provides information on which students need additional support or are at risk of dropping out. With a view to encourage and motivate those students, remedial actions could be organized, such as early alerts and advising interventions.In real-world problems, it is more of a rule than an exception that the available data may contain inaccurate data, are incomplete and/or are appeal to subjectivity. It is also often necessary the exported learning models to have an easy-to-understand representation and transparency of how they work, in order to forge a synergy between humans and machines. At the same time, the rapid development and the wide variety of advanced Machine Learning solutions require specialized knowledge and experience in the relevant fields from their users, thus preventing the participation of inexperienced ones.This dissertation deals with the general problem of creating predictive classification and regression models in the field of Education. The main objective is to develop methodologies that efficiently construct reliable learning models for predicting the performance of students in the final examinations of semester courses of higher education.To this end, six new research studies are described which provide new methods of extracting knowledge from educational data. Regardless of their individual differences, the studies were based upon two basic principles: the handling of incomplete information that generally exists in digital data (educational and not) and the employment of techniques that produce high-performance models with intuitively interpretable structure, without excluding non-specialists in the field due to the complexity of the available options that usually surround them.In this context, Fuzzy Learning Algorithms, Automated Machine Learning techniques and Active Learning strategies were employed and progressively combined, while the Transfer Learning technique was separately studied. In a first step, the effectiveness of Bayesian optimization in the automatic selection of the optimal learning algorithm and in the adjustment of their hyper-parameters was investigated. To equally highlight the importance of producing interpretable and explainable Machine Learning models, the search space was restricted to tree-based and rule-based algorithms. The result of this consolidation was the formation of models with human-understandable representation, which were also characterized by high accuracy.In addition, the use of Active Learning and Transfer Learning strategies is proposed. These techniques are focused on addressing the difficulty of collecting sufficient data that are required to fit a learning model. Active Learning is well-motivated in problems, where unlabeled data may be abundant or easy to come by, but training labels are difficult, or expensive to obtain. Transfer Learning is well-motivated in problems, where there is a lack of data regarding a problem, but there is plenty of data about another related one. Two distinct and innovative methodologies have been developed: a hybrid Active Learning method based on Fuzzy classifiers that incorporates automated optimization mechanisms and a Transfer Learning method between predictive learning models. The new proposed methods are applied to classification and regression problems using mainly educational data describing the student activity on the Moodle online learning platform. The experimental results document the effectiveness of these methods for the problem of predicting students' learning outcomes.Hopefully, the systematic research that has been conducted could be considered as a useful contribution in the fields of Educational Data Mining and Learning Analytics. At the same time, it could lead to the development of tools that assist instructors and learning through data-driven decisions.

Degree

thesis:*
Grantor dc:publisher
University of Patras
Year dc:date
2021

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Tsiakmaki, Maria
  • Τσιακμάκη, Μαρία

Subjects

dc:subject × 18

Rights

Language dc:language
gre

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/51350

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

Tsiakmaki, Maria; Τσιακμάκη, Μαρία. Νέοι αλγόριθμοι μηχανικής μάθησης για την επαγωγή γνώσης από εκπαιδευτικά / μαθησιακά δεδομένα. University of Patras, 2021. http://hdl.handle.net/10442/hedi/51350