Improved clustering and classification algorithms for the Kohonen self-organizing neural network
Mohamed Abdalla Nour · 1994
The Kohonen Self-Organizing Neural Network possesses remarkable properties such as computational simplicity, use of unsupervised learning, and applicability to a broad class of problems. Coupled with these advantages, however, are some serious limitations that handicap its learning algorithm from achieving superior performance, especially with regard to pattern clustering and classification. Some of these limitations are: (1) slow convergence, (2) computational inefficiencies, and (3) suboptimal results. There have been extensive studies over the last decade to address these limitations by extending the Kohonen self-organizing learning algorithm using heuristic and optimization approaches. Unfortunately, the extent of improvements resulting from these research efforts have been relatively insignificant, while most of the extensions are not applicable to pattern clustering and classification, an interesting area in many business applications. This study develops several extensions to the standard Kohonen self-organizing algorithm. We first critically analyze previous extension efforts with a view of delineating opportunities for performance improvements. Two improved versions of the standard Kohonen learning algorithm are then proposed, tested, and compared with the standard version. Both modified versions are based on a new network architecture that serves as a point of departure from the standard Kohonen self-organizing model. These algorithms are naturally suited to applications in pattern clustering and classification. In all the computer simulations using several test data sets, the modified versions have shown superior performance to the standard version. We have also tested the three algorithms on two financial applications, corporate bond rating and bankruptcy prediction, and compared their performance results with those from conventional statistical methods, namely regression analysis and discriminant analysis. Our computer simulations have indicated comparable classification performances from the neural network algorithms and the two statistical methods.