| Abstract [eng] |
Often in the context of multidimensional data, there is a need to analyze the clusters, find similarities or differences between the data and to deal with various challenges of the classification. Increasingly, we are surrounded by a various text data, so it is necessary to find effective methods to analyze these kind of data. The area of research is analysis of the text and numeric datasets by using the self-organizing maps (SOM). In the dissertation, the biggest focus is on the text data visualization and evaluation of the quality of the resulting map, using self-organizing maps. In this work, there is proposed and implemented the SOM visualization way, which helps a researcher to see the different classes of data in the same self-organizing map cell. Also, the two new errors are proposed, which are helpful to estimate the SOM map quality. The first error shows how close the same class members in the SOM are. The smaller value of error means the better results, it means that all same class members are closer to each other, the clusters are “stronger”. The second error shows how far the centers of different classes are. The bigger value of error means the better results, i.e. all the different class centers are far from each other, so they are separated on the map. Both errors are suitable when classified data are analyzed. In this work, the new SOM system is developed, in which we can use the purposed SOM visualization way and two new errors. A comparative analysis of the most popular SOM systems and the new system have been done. The dependence of the self-organizing maps learning parameters to SOM results has been defined experimental. Also, how different factors of text data conversation to numerical expression make influence to SOM results are experimentally investigated. |
