Abstract
dc:descriptionRecognizing unimodal data distributions is of great significance in statistics, machine learning and data science. The characteristic property of a unimodal distribution is that data values are gathered around a single value (peak), which is the mode of the distribution. Due to this property, data can be characterized as homogeneous, forming a single and coherent group. Well-known distributions, such as Gaussian, Student’s t and Gamma are typical examples of unimodal distributions. Also, the uniform distribution is considered as an extreme unimodal case. Unimodality tests have been proposed to decide on the unimodality of a set of data values, thus providing useful knowledge about the structure of the data. This thesis concerns the development and implementation of machine learning methods based on the notion of unimodality, focusing on four main axes: i) the creation of a new unimodality test for deciding data unimodality, ii) the analysis of key characteristics of data density, such as modes and valleys, which leads to the discovery of innovative properties explored in detail, iii) the development of statistical models, specifically mixture models, for modeling univariate unimodal and multimodal (multiple peaks) data, and iv) the development of partitioning methods for multidimensional data into clusters that are unimodal along each axis, achieved through the unsupervised construction of axis-aligned binary decision trees. We begin, by proposing a new unimodality test called Unimodal Uniform test (UU-test) to decide if a dataset has been generated by a unimodal distribution or not. The method utilizes the empirical distribution function (ecdf) and attempts to obtain a unimodal piecewise linear approximation of the ecdf under the constraint that the data corresponding to each linear segment follow the uniform distribution. Compared to other unimodality tests, it also produces a generative model of the unimodal data in the form of a mixture of uniform distributions (UMM). Thus, it can be used for statistical data modeling of unimodal distributions with arbitrary shape. Next, we improve UMM performance by substituting the uniform distribution with a more flexible and differential one, called Π-sigmoid. The Π-sigmoid distribution, defined as the difference of two translated logistic sigmoids, can approximate a wide range of distributions. We employ and train a mixture model of Π-sigmoids, called UΠsMM, initialized using the output of the UU-test. Additionally, we introduce a mechanism to maintain the unimodality of the model during training via the Expectation-Maximization (EM) algorithm. UΠsMM achieves an accurate fit while often requiring fewer components than UMM. Afterward, we address the problem of modeling univariate multimodal data, with two main contributions. First, we introduce properties of critical points of the data ecdf that provide indications on the existence of density valleys. Using these properties, we propose UniSplit, an algorithm that detects valley points and partitions the dataset into unimodal subsets, automatically estimating their number. Second, we propose a statistical model, the Unimodal Mixture Model (UDMM), which models each unimodal subset with a UMM. A key strength of UDMM is its flexibility and independence from specific parametric assumptions, making it well-suited for datasets generated by sources of different probability density (e.g., one Gaussian and one uniform). Another important property is that the number of components is automatically estimated, therefore, a major issue in mixture modeling is addressed. Finally, we focus on developing an unsupervised method for clustering multidimensional data using decision trees. We introduce the concept of axis unimodal clusters, i.e., clusters where all features are unimodal as decided by a unimodality test. We present a method that constructs binary decision trees, providing axis-aligned partitions of the data and offering interpretable clustering solutions. Two criteria are proposed to identify the best split pair (feature and threshold) at each node, aiming to improve the unimodality of the partition after each split. Compared to other unsupervised decision tree methods, this approach has several advantages: it is simple, avoids preprocessing steps and does not employ computationally expensive optimization methods or difficult to tune hyperparameters, such as number of clusters or maximum tree depth.
Degree
thesis:*- Grantor dc:publisher
- University of Ioannina
- Year dc:date
- 2025
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Chasani, Paraskevi
- Χασάνη, Παρασκευή
Subjects
dc:subject × 14Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/58427