University of the Aegean
Recent advances on dimensionality reduction for high-dimensional data analysis with applications
Abstract
dc:descriptionLarge amounts of raw data often can fail to perform properly for model estimation, attributed to the existence of multicollinearity between variables, and that is why they must be pre-processed for better modeling and visualization. To address raw data barriers, among other difficulties, Dimension Reduction Techniques were developed in an effort to mitigate the magnitude of over-parametrized solutions that arise in high-dimensional spaces. The aim of this dissertation, which utilizes multivariate analysis tools, is to investigate, analyze, compare, and improve current techniques while still introducing new ones for dealing with multicollinearity and reducing the feature space of high-dimensional data. In particular, this doctoral thesis initially outlines the theoretical framework concerning the unsupervised technique, Principal Component Analysis, and its supervised counterpart, the Partial Least Squares method. Due to their ability to obtain dimension reduction when analyzing high-dimensional datasets, both techniques are considered optimal for feature extraction. The use of the former in conjunction with other dimension reduction techniques, as well as the modification of the latter, - so that it may be applied as a feature selection and feature extraction simultaneously-, were implemented and thoroughly studied in the fields of econometrics, finance and actuarial science. Finally, a new unsupervised linear feature selection technique is proposed as a robust and easily interpretable methodology, termed Elastic Information Criterion, that is capable of capturing multicollinearity rather accurately and effectively and thus providing a proper dataset assessment.
Degree
thesis:*- Grantor dc:publisher
- University of the Aegean
- Year dc:date
- 2022
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Ntotsis, Kimon
- Ντότσης, Κίμων
Subjects
dc:subject × 22- Τεχνικές μείωσης διαστάσεων
- ΠΟΛΥΣΥΓΓΡΑΜΜΙΚΟΤΗΤΑ
- Ανάλυση κύριων συνιστωσών (ΑΚΣ)
- Μέθοδος μερικών ελαχίστων τετραγώνων
- Μέθοδος ποινικοποίησης
- Μέθοδος επιλογής μεταβλητών
- Μέθοδος επιλογής μοντέλου
- Dimension reduction techniques
- MULTICOLLINEARITY
- Principal component analysis (PCA)
- Partial Least Squares Regression
- Regularization methods
- Variable selection methods
- Model selection
- Μαθηματικά
- Φυσικές Επιστήμες
- Στατιστική και Πιθανότητες
- Μοντελοποίηση και Προσομοίωση
- Mathematics
- Natural Sciences
- Statistics and Probability
- Modeling and Simulation
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/52496