The Development and Application of a Three-Layered Self-Organizing Map Neural Network

• Main Authors: Chi-Chung Li, 李紀忠
• Other Authors: Sheng-Chai Chi
• Format: Others
• Language: zh-TW
• Published: 2002
• Online Access: http://ndltd.ncl.edu.tw/handle/83712634870771736259

碩士 === 華梵大學 === 工業管理學系碩士班 === 90 === In today’s commercial activities, holding the effective information through information technology and the internet becomes a very important symptom to judge if the enterprise has business competitive strengths. It used to be not so easy to collect information in the past time. Also, it took a long time and spent a high cost to get some part of useful data. With the continuous improvement of information technology nowadays, how to mine out useful information from a huge database for problem solving or making correct decisions by enterprise or individual has become an important research issue. Clustering analysis, a technique for data mining or data analysis from a database, has been widely applied in various areas. Its purpose is to segment the individuals, which have the same characteristics in the population, into the same group. Thus, those belonged to the same group are homogeneous and have the same characteristics; those belonged to different groups are heterogeneous and have different characteristics. In this research, we attempt to enhance the conventional two-layered self-organizing map (SOM) into three-layered SOM neural network for the avoidance of human judgment after the network topology has been mapped and study the performance of the proposed SOM neural network in clustering analysis. In developing the three-layered SOM, two training modes (continuous training mode and two-staged training mode) and two data manipulating methods for preparing the input of the second layer are developed. In order to compare these four models of the SOM, a set of part images with a given result is applied to evaluate their performance in clustering analysis. In addition, this research also tests the clustering performance of the proposed SOM with the involvement of different percentages of input noises. In this research, the enhanced three-layered SOM is developed for overcoming the drawback of further identification of the clusters on the network topology required by the conventional two-layered SOM through sight-inspection or other methods. From the results of the experiments, it has been proven that the data belonged to the same group can be mapped to the same neuron on the output layer of the three-layered SOM and the accuracy is very high.