University of Patras
Προηγμένες τεχνικές μηχανικής μάθησης για πρόβλεψη και ταξινόμηση με χρήση συναισθηματικής ανάλυσης και ενεργής μάθησης
Abstract
dc:descriptionThis doctoral dissertation focuses on exploring Machine Learning techniques that incomporate Sentiment Analysis and Active Learning, with the primary goal of enhancing existing, efficient methods. The issues addressed are grouped into two main sections. In the first section, we concentrate on the use of Sentiment Analysis in financial domain applications. Initially, we investigate the scenario of stock market forecasting by utilizing data from social networks—specifically Twitter—to capture and assess public opinion. To this end, two extensive experimental procedures were designed: the first relies primarily on classical Machine Learning methods, whereas the second employs Deep Learning techniques. In both cases, we examine not only the performance of multiple algorithms but also the effect of different sentiment configurations (TextBlob, VADER, FinBERT). The findings highlight the positive contribution of Sentiment Analysis to improving economic forecasts. Subsequently, the study focuses on Financial Sentiment Analysis as a classification problem, comparing on the one hand numerous classical methods and their ensembles, and on the other hand various BERT variants. Concurrently, the effectiveness of class-balancing techniques (e.g., SMOTE) is examined in combination with different language representations (BoW, TF-IDF). The results highlight the superiority of BERT-based models—particularly RoBERTa—and show that the SMOTE and TF-IDF combination further optimizes performance. The second section addresses the use of Active Learning to reduce the requirement for labeled data in classification tasks. Two distinct scenarios were implemented. In the first, Active Learning is applied to the problem of Software Defect Prediction, employing both individual Machine Learning methods and their ensemble versions. In the second, we investigate whether comparable performance across various metrics can be achieved when Sentiment Analysis is supported by Transformer models that incorporate Active Learning. In both scenarios, the targeted selection of examples to be labeled yields models with performance analogous to that obtained by training on the entire dataset. Finally, a “meta-index” for evaluating Active Learning scenarios is proposed, suited to related experimental contexts, with the goal of providing a more comprehensive assessment of the effectiveness of the proposed methods.
Degree
thesis:*- Grantor dc:publisher
- University of Patras
- Year dc:date
- 2025
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Karanikola, Aikaterini
- Καρανικόλα, Αικατερίνη
Subjects
dc:subject × 20- Μηχανική μάθηση
- Ανάλυση συναισθήματος
- Ενεργή μάθηση
- Πρόβλεψη χρονοσειρών
- Κατηγοριοποίηση κειμένου
- Machine learning
- Sentiment analysis
- Active learning
- Time series forecasting
- Text classification
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Φυσικές Επιστήμες
- Τεχνητή νοημοσύνη
- Μαθηματικά
- Υπολογιστικά μαθηματικά
- Computer and Information Sciences
- Natural Sciences
- Artificial Intelligence
- Mathematics
- Computational Mathematics
Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/59690