Data exploration process based on the self-organizing map

Abstract

With the advances in computer technology, the amount of data that is obtained from various sources and stored in electronic media is growing at exponential rates. Data mining is a research area which answers to the challange of analysing this data in order to find useful information contained therein. The Self-Organizing Map (SOM) is one of the methods used in data mining. It quantizes the training data into a representative set of prototype vectors and maps them on a low-dimensional grid. The SOM is a prominent tool in the initial exploratory phase in data mining.

The thesis consists of an introduction and ten publications. In the publications, the validity of SOM-based data exploration methods has been investigated and various enhancements to them have been proposed. In the introduction, these methods are presented as parts of the data mining process, and they are compared with other data exploration methods with similar aims.

The work makes two primary contributions. Firstly, it has been shown that the SOM provides a versatile platform on top of which various data exploration methods can be efficiently constructed. New methods and measures for visualization of data, clustering, cluster characterization, and quantization have been proposed. The SOM algorithm and the proposed methods and measures have been implemented as a set of Matlab routines in the SOM Toolbox software library.

Secondly, a framework for SOM-based data exploration of table-format data - both single tables and hierarchically organized tables - has been constructed. The framework divides exploratory data analysis into several sub-tasks, most notably the analysis of samples and the analysis of variables. The analysis methods are applied autonomously and their results are provided in a report describing the most important properties of the data manifold. In such a framework, the attention of the data miner can be directed more towards the actual data exploration task, rather than on the application of the analysis methods. Because of the highly iterative nature of the data exploration, the automation of routine analysis tasks can reduce the time needed by the data exploration process considerably.