Aristotle University Of Thessaloniki (AUTH)
Μείωση μαθηματικών διαστάσεων πολυμεταβλητών κατηγορικών δεδομένων και μηχανική μάθηση: νέες μεθοδολογικές προτάσεις και επεκτάσεις
Abstract
dc:descriptionThis dissertation investigates the relationship and complementarity between Data Analysis and Machine Learning, with particular emphasis on high-dimensional categorical data. Such data are widely encountered in scientific fields, including biological sciences, social sciences, and education, yet they pose significant analytical challenges due to their non-numerical nature and the complexity of the relationships among variables. The primary objective of this dissertation is to investigate, compare, and extend dimensionality reduction strategies for categorical data in combination with Machine Learning classification algorithms. The integration of these approaches enables a more comprehensive exploration of data structure and variability, facilitating the extraction of meaningful information and the production of more reliable and interpretable results. A hybrid methodological framework is adopted, combining classical multivariate statistical techniques with modern Machine Learning algorithms. The study examines several dimensionality reduction methods, including Multiple Correspondence Analysis (MCA), Categorical Principal Component Analysis (CATPCA), Principal Component Analysis (PCA) following appropriate quantification of categorical variables, Factor Analysis of Mixed Data (FAMD), as well as Nonlinear Canonical Correlation Analysis (NLCCA) and Multiple Factor Analysis (MFA). Alternative dimensionality reduction strategies are evaluated based on the variable’s measurement scale (nominal, ordinal, or mixed), using appropriate transformations, optimal scaling procedures, and combinations of these techniques. The effectiveness of the proposed dimensionality reduction strategies is further assessed using supervised Machine Learning classification algorithms, including Random Forests, Artificial Neural Networks, k-Nearest Neighbors (k-NN), Support Vector Machines (SVM), and Logistic Regression. In addition, feature selection is examined as an alternative to dimensionality reduction to improve classification performance. A major contribution of this dissertation is the development of three methodological extensions of Correspondence Analysis (CA/MCA). The first extends the application of CA/MCA and CA-raw from contingency tables to quantitative data matrices by transforming quantitative values into integer frequencies, thereby enabling analysis of secondary quantitative data and broadening their applicability. The second applies CA/MCA to matrices in which rows represent complete category interaction profiles rather than individual observations. This approach achieves substantial information compression, improves the interpretation of distances in the factorial space, and more effectively reveals interactions among categorical variables. The third combines Correspondence Analysis with Kriging spatial interpolation and Hierarchical Cluster Analysis (HCA), providing an integrated analytical framework that links geometric representation, clustering, and spatial visualization for the analysis of complex compositional data. The empirical evaluation is based on real-world datasets at both small and large scales, enabling a comprehensive comparison of alternative dimensionality reduction strategies. The results demonstrate that the choice of dimensionality reduction method has a significant impact on both predictive accuracy and model interpretability. Furthermore, the findings indicate that no universally optimal dimensionality reduction or classification method exists; instead, the most appropriate approach depends on the research objectives and the characteristics of the data. Overall, this dissertation contributes to both the theoretical understanding of the relationship between Data Analysis and Machine Learning and the methodological advancement of high-dimensional categorical data analysis. The proposed approaches broaden the applicability of existing methods and provide a foundation for future research on developing new and alternative strategies for the analysis and management of complex categorical data.
Degree
thesis:*- Grantor dc:publisher
- Aristotle University Of Thessaloniki (AUTH)
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Papafilippou, Nikolaos
- Παπαφιλίππου, Νικόλαος
Subjects
dc:subject × 30- Πολυδιάστατα κατηγορικά δεδομένα
- Στρατηγικές μείωσης διαστάσεων
- Ανάλυση δεδομένων
- Μηχανική μάθηση
- Οπτικοποίηση πολυμεταβλητών δεδομένων
- Biplot ανάλυση
- Eπιλογή χαρακτηριστικών
- Προβλεπτική ταξινόμηση
- Εναλλακτικές στρατηγικές ανάλυσης
- Πίνακας πλήρων συνδυασμών κατηγοριών
- Multidimensional categorical data
- Dimensionality reduction strategies
- Data analysis
- Machine learning
- Multivariate data visualization
- Biplot analysis
- Feature selection
- Predictive classification
- Alternative analysis strategies
- Full category combination table
- Μαθηματικά
- Φυσικές Επιστήμες
- Στατιστική και Πιθανότητες
- Άλλες Γεωπονικές Επιστήμες
- Γεωπονικές Επιστήμες και Κτηνιατρική
- Mathematics
- Natural Sciences
- Statistics and Probability
- Other Agricultural Sciences
- Agricultural and Veterinary Sciences
Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/62475