Development of a New Means to Improve the Performance of Self-Organizing Maps

Abstract

In machine learning, a Self-Organizing map plays a significant role in finding hidden patterns or intrinsic structures in data. In this study, a new modified expression is arrived at to update the radius of the neighbourhood of BMU in SOM. Further, a new approach is introduced to find the eligible nodes for an update in SOM. We have also incorporated the previous work, such as new initialization algorithms to find the initial weight vectors, a method to place the weight vectors in each node of the grid and a method to identify the number of clusters to improve the performance of the proposed SOM algorithm. The proposed SOM performance in terms of Quantization error (QE), Convergence time (CT) and Modified Semantic Relevance Index (MSRI) are compared with conventional SOM for both class label and non-class label datasets. In addition to the above measure, Classification Accuracy (CA) is also used to evaluate the performance against class label datasets. The proposed SOM algorithm shows better performance in all cases.