Tastan, Aylin (2023)
Contributions to Robust Graph Clustering: Spectral Analysis and Algorithms.
Technische Universität Darmstadt
doi: 10.26083/tuprints-00024068
Ph.D. Thesis, Primary publication, Publisher's Version
Text
2023-05-31_Tastan_Aylin.pdf Copyright Information: CC BY-NC-SA 4.0 International - Creative Commons, Attribution NonCommercial, ShareAlike. Download (30MB) |
Item Type: | Ph.D. Thesis | ||||
---|---|---|---|---|---|
Type of entry: | Primary publication | ||||
Title: | Contributions to Robust Graph Clustering: Spectral Analysis and Algorithms | ||||
Language: | English | ||||
Referees: | Zoubir, Prof. Dr. Abdelhak M. ; Muma, Prof. Dr. Michael | ||||
Date: | 2023 | ||||
Place of Publication: | Darmstadt | ||||
Collation: | xii, 268 Seiten | ||||
Date of oral examination: | 26 May 2023 | ||||
DOI: | 10.26083/tuprints-00024068 | ||||
Abstract: | This dissertation details the design of fast, and parameter free, graph clustering methods to robustly determine set cluster assignments. It provides spectral analysis as well as algorithms that adapt the obtained theoretical results to the implementation of robust graph clustering techniques. Sparsity is of importance in graph clustering and a first contribution of the thesis is the definition of a sparse graph model consistent with the graph clustering objectives. This model is based on an advantageous property, arising from a block diagonal representation, of a matrix that promotes the density of connections within clusters and sparsity between them. Spectral analysis of the sparse graph model including the eigen-decomposition of the Laplacian matrix is conducted. The analysis of the Laplacian matrix is simplified by defining a vector that carries all the relevant information that is contained in the Laplacian matrix. The obtained spectral properties of sparse graphs are adapted to sparsity-aware clustering based on two methods that formulate the determination of the sparsity level as approximations to spectral properties of the sparse graph models. A second contribution of this thesis is to analyze the effects of outliers on graph clustering and to propose algorithms that address robustness and the level of sparsity jointly. The basis for this contribution is to specify fundamental outlier types that occur in the cases of extreme sparsity and the mathematical analysis of their effects on sparse graphs to develop graph clustering algorithms that are robust against the investigated outlier effects. Based on the obtained results, two different robust and sparsity-aware affinity matrix construction methods are proposed. Motivated by the outliers’ effects on eigenvectors, a robust Fiedler vector estimation and a robust spectral clustering methods are proposed. Finally, an outlier detection algorithm that is built upon the vertex degree is proposed and applied to gait analysis. The results of this thesis demonstrate the importance of jointly addressing robustness and the level of sparsity for graph clustering algorithms. Additionally, simplified Laplacian matrix analysis provides promising results to design graph construction methods that may be computed efficiently through the optimization in a vector space instead of the usually used matrix space. |
||||
Alternative Abstract: |
|
||||
Status: | Publisher's Version | ||||
URN: | urn:nbn:de:tuda-tuprints-240689 | ||||
Classification DDC: | 500 Science and mathematics > 510 Mathematics 600 Technology, medicine, applied sciences > 620 Engineering and machine engineering |
||||
Divisions: | 18 Department of Electrical Engineering and Information Technology > Institute for Telecommunications > Robust Data Science 18 Department of Electrical Engineering and Information Technology > Institute for Telecommunications > Signal Processing |
||||
Date Deposited: | 07 Jun 2023 07:17 | ||||
Last Modified: | 09 Jun 2023 06:21 | ||||
URI: | https://tuprints.ulb.tu-darmstadt.de/id/eprint/24068 | ||||
PPN: | 508366704 | ||||
Export: |
View Item |