Aristotle University Of Thessaloniki (AUTH)
Σύγκριση μεθόδων μείωσης διαστάσεων και ταξινόμησης σε «Κύβους Δεδομένων»: νέες μεθοδολογικές προτάσεις και επεκτάσεις
Abstract
dc:descriptionMultidimensional Data Analysis is a "cornerstone" of Data Analysis. In this context, data cubes are p-dimensional data structures, where p≥3. This thesis focuses exclusively on data cubes with independent "slices". Understanding the structure and management of cubes is a springboard for the development of alternative methodological approaches to small and big data analysis, beyond the classical methods of dimensionality reduction and clustering. Therefore, the main objective of this doctoral thesis is to propose new ideas related to the development of data cube dimensionality reduction and clustering strategies, providing researchers with important alternative tools for data analysis. At the same time, it seeks to highlight the importance of applying different dimensionality reduction and clustering strategies for data cubes, as well as different ways of managing a dataset, based on its structure and the objectives of each study. There are two basic methodological approaches for analyzing data cubes. The first is based on analyzing the data cube as a two-dimensional mixed-type dataset, in which the "slices" are considered additional categorical variables. The second is based on analyzing the data cube as a "stacked" data matrix (after vertical concatenation), in which the "slices" and "hyper-slices" do not lose their conceptual and structural role. Within the framework of the first methodological approach, the classic methods Principal Components Analysis (PCA), Factor Analysis, Multiple Correspondence Analysis, Non-Linear Categorical Principal Components Analysis with optimal scaling, and Factor Analysis for Mixed Data are applied and compared under different combinations of data encoding and conversion systems, as well as criteria for extracting the most important factorial axes. In addition, a comparison of the statistical software IBM SPSS Statistics, R, and Python is carried out in terms of the results extraction times. Within the framework of the second methodological approach, 13 dimensionality reduction strategies are applied and compared on three-dimensional cubes, based on PCA, which is an important Data Analysis method and pre-processing stage of Machine Learning. Some strategies are based on the idea of decomposing the total variability of the cube into variability between-"slices" and variability within-"slices" of the cube. Some other strategies either initially ignore the effect of "slices" on the structure and components of variability and introduce it at a second level, or completely ignore the effect of "slices" on the structure and components of variability, or take into account the effect of "slices" on the structure and components of variability through different coding of the "slices" variable. Within the framework of the second methodological approach, an alternative dimensionality reduction strategy on four-dimensional cubes is also proposed and implemented, based on PCA. The strategy is based on the idea of decomposing the total variability of the cube into variability between-"slices", variability between-"hyper-slices", interaction variability between "slices" and "hyper-slices", and variability within "slices" and "hyper-slices". In addition to the above methodological approaches, clustering strategies are proposed that vary in terms of the chosen distance matrix and linkage method. To implement all the proposed strategies, the real datasets "Forest Cover Type", "Iris Plants", "ADONUT" and "Basil" from the broader field of biological sciences, as well as the larger simulated datasets "Basil-Sim 1" and "Basil-Sim 2" are used. Python code for creating the simulated datasets, as well as code for implementing the proposed analysis strategies in IBM SPSS Statistics, R, and Python is provided. In conclusion, the comparisons highlight the advantages and disadvantages of the proposed strategies, as well as the advantages and computational weaknesses and difficulties of statistical software. In addition, the contribution or non-contribution of "slices" and/or "hyper-slices" in the structure of the data cubes is highlighted, as well as the fact that there is not a "global" method of dimensionality reduction and clustering for the objectives of each study. Therefore, it is considered necessary to apply many different analysis strategies in order to increase the range of applications of each method and to extract more useful results. The contribution of the proposed strategies to the scientific community is considered significant because they constitute alternative tools for data management and analysis, which can be adapted according to the objectives of each study and the observed data. This thesis aims to serve as a springboard for further research and development of new and/or alternative data analysis and management strategies.
Degree
thesis:*- Grantor dc:publisher
- Aristotle University Of Thessaloniki (AUTH)
- Year dc:date
- 2025
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Kyrana, Zacharenia
- Κυρανά, Ζαχαρένια
Subjects
dc:subject × 43- Ανάλυση δεδομένων
- Κύβοι Δεδομένων
- Πολυδιάστατη ανάλυση δεδομένων
- Πολυμεταβλητή ανάλυση δεδομένων
- Βιομετρία
- Διαμέριση μεταβλητότητας κύβων
- Προσομοίωση δεδομένων
- Μείωση διαστάσεων
- Ταξινόμηση-συσταδοποίηση
- Ανάλυση σε κύριες συνιστώσες
- IBM SPSS Statistics
- Python
- R
- «Φέτες»
- «Υπερ-φέτες»
- Κατακόρυφη «συνένωση»
- Πίνακας «στοίβα»
- Στρατηγικές
- Data analysis
- Data Cubes
- Multidimensional data analysis
- Multivariate data analysis
- Biometry
- Cubes variability decomposition
- Data simulation
- Dimensionality reduction
- Clustering
- Principal components analysis (PCA)
- "Slices"
- "Hyper-slices"
- Vertical concatenation
- Stacked matrix
- Strategies
- 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/60622