Back to search

University of Ioannina

Clustering methods based on deep learning and unimodality testing

Abstract

dc:description

Data clustering is the process of partitioning a dataset into a finite set of groups, or clusters, such that data points within each cluster exhibit intra-cluster similarity, while those belonging to different clusters are characterized by inter-cluster dissimilarity. Clustering remains a challenging task due to the inherent complexity of uncovering meaningful structures within data. Revealing these hidden structures provides valuable insights and facilitates a deeper understanding of the underlying patterns. This thesis concerns the development, implementation and evaluation of novel clustering methodologies mainly focused on three important problems: i) partitional clustering in both Euclidean and kernel spaces, ii) unimodality-based clustering, which incorporates the concept of unimodality into the clustering process, and iii) deep clustering, which leverages the representational power of deep learning methods. We first introduce global k-means++, a method developed to address the initialization challenges inherent in the standard k-means algorithm. The approach integrates the incremental strategy of global k-means with the probabilistic center selection mechanism of k-means++, effectively combining the strengths of both techniques. The resulting synergy delivers high-quality clustering solutions while significantly reducing the computational cost typically associated with global k-means. Furthermore, we extend this concept from Euclidean to kernel space by proposing global kernel k-means++, an algorithm specifically designed to overcome the initialization problem in kernel k-means. The optimization effectiveness of both global k-means variants is thoroughly validated through extensive experimental evaluation. Afterwards, we present UniForCE, a clustering method that simultaneously partitions data and estimates the number of clusters k. UniForCE introduces a novel notion of locally unimodal clusters, focusing on unimodality at local regions of the data density rather than in the entire cluster. By identifying unimodal pairs of neighboring subclusters, the method aggregates them into larger, statistically coherent structures via a unimodality graph. This flexible formulation enables the discovery of arbitrarily shaped clusters. A statistical test determines unimodal pairs, and clustering is achieved with automatic estimation k by detecting the number of connected components in the unimodality graph. Extensive experiments on synthetic and real datasets validate both the conceptual soundness of the method and its practical effectiveness. Furthermore, we introduce the soft silhouette score, a generalization of the widely used silhouette measure that accommodates probabilistic cluster assignments. Building on this differentiable measure, we develop an autoencoder-based deep clustering method utilizing the soft silhouette score. Our method guides the learned latent representations to form clusters that are both compact and well-separated. This property is crucial in real-world applications, as simultaneously ensuring compactness and separability guarantees that clusters are not only densely packed but also clearly distinct from each other. We evaluate our method on a variety of benchmark datasets and against state-of-the-art methods to demonstrate that it outperforms established deep clustering approaches, highlighting the effectiveness of the soft silhouette score as a principled objective for improving the quality of learned latent representations. Finally, we present the neural implicit maximum likelihood clustering, which is a neural-network-based approach that frames clustering as a generative task within the Implicit Maximum Likelihood Estimation framework. By adapting ideas from ClusterGAN, our method avoids several well-known shortcomings of GAN-based clustering while maintaining a simple and stable training objective. The method performs particularly well on small datasets, with experimental comparisons against both deep and conventional clustering algorithms underscoring its competitive potential. A notable strength of our method is its ability to capture diverse cluster geometries without requiring hyperparameter tuning. Experiments on synthetic datasets show that the method can successfully cluster both cloud-shaped and ring-shaped data.

Degree

thesis:*
Grantor dc:publisher
University of Ioannina
Year dc:date
2025

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Vardakas, Georgios
  • Βαρδάκας, Γεώργιος

Subjects

dc:subject × 13

Rights

Language dc:language
eng

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/60102

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Vardakas, Georgios; Βαρδάκας, Γεώργιος. Clustering methods based on deep learning and unimodality testing. University of Ioannina, 2025. http://hdl.handle.net/10442/hedi/60102