Abstract
dc:descriptionIn this PhD thesis, several signal-processing models applied on graphs have been studied. Half-quadratic (HQ) optimization, robust estimators, such as marginal median and M-estimators, l21 norm regularization, block least mean squares have been the fundamental basis of the proposed approaches. Specifically, to remedy the vulnerability from erroneous embeddings on Multidimensional Scaling (MDS) problem due to the distortion introduced by outliers, a unified framework is proposed where the MDS is treated as maximization of a correntropy criterion, which is solved by HQ optimization in either multiplicative or additive forms. Three novel algorithms are derived. Their performance is assessed experimentally against three state-of-the-art MDS techniques, namely the Scaling by Majorizing a Complicated Function, the Robust Euclidean Embedding, and the Robust MDS under the same conditions. The experimental results indicate that the proposed algorithms perform substantially better than the aforementioned competing techniques. Furthermore, another unified framework for the solution of MDS problem is proposed, which resorts, again, to HQ optimization and employs potential functions of M-estimators in combination with l21 norm regularization. The numerical tests demonstrate that the proposed algorithms perform better than the competing alternatives. Moreover, a novel HQ framework, employing M-estimators and nuclear norm regularization, has been proposed for solving the MDS problem in the presence of a dissimilarity matrix corrupted with outliers. It has been proven that the proposed algorithms outperform the state-of-the-art MDS techniques. Then, the inclusion of several dimensionality reduction (DR) algorithms within the unified framework, known as graph embedding, is demonstrated. The enforcement of orthogonality to the projection matrix within this framework is proven to be of vital importance, since the use of orthogonal mapping DR algorithms against their non-orthogonal counterparts improves significantly the recognition rate of Nearest Neighbor (NN) classifier. One may simultaneously also impose the l21 regularization on the projection matrix, seeking for row-sparsity. This leads to the known Joint Feature Selection and Subspace Learning (JFSSL) framework. All DR algorithms are employed within JFSSL. It is proven that the use of orthogonal mapping algorithms within JFSSL against their non-orthogonal counterparts does not improve the recognition rate of NN classifier. Then, the case when a randomly chosen subset of agents shares deliberately with its neighborhood local information corrupted with gross errors is considered. A three-fold extension of the state-of-the-art spatio-temporal diffusion algorithms over mobile networks is proposed. First, the agents preserve a limited memory of prior regression vectors and prior estimates of the distance to the target in the time domain so that a block least mean squares is applied. Second, in the time domain, the agents estimate the combination weights employing the M-estimators’ weighting function of the scaled distance of past intermediate location estimates from their mean. Third, in the spatial domain, the agents estimate the combination weights based on negative exponentials of the scaled distance of intermediate location estimates within their neighborhood from the marginal median of these estimates. The experiments demonstrate that, at the expense of some additional computational requirements, the proposed algorithms can cope with the outliers and outperform the contemporary strategies.
Degree
thesis:*- Grantor dc:publisher
- Aristotle University Of Thessaloniki (AUTH)
- Year dc:date
- 2023
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Mandanas, Fotios
- Μανδάνας, Φώτιος
Subjects
dc:subject × 16- Τετραγωνική κατά το ήμισυ ελαχιστοποίηση
- Εύρωστοι εκτιμητές
- Κανονικοποίηση μέσω l21 νόρμας
- Ομαδική μέθοδος ελαχίστων τετραγώνων
- Half-quadratic minimization
- ROBUST ESTIMATORS
- l21 norm regularization
- Block least mean squares
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Φυσικές Επιστήμες
- Επιστήμη ηλεκτρονικών υπολογιστών, θεωρία και μέθοδοι
- Επιστήμη ηλεκτρονικών υπολογιστών και Πληροφορική, άλλοι τομείς
- Computer and Information Sciences
- Natural Sciences
- Computer science, theory and methods
- Computer and Information sciences, miscellaneous
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/53514