A Comparative Performance Analysis of Self Organizing Maps on Weight Initializations Using different Strategies

Self Organizing Maps perform clustering of data based on unsupervised learning. It is of concern that initialization of the weight vector contributes significantly to the performance of SOM and since real world datasets being high-dimensional, the complexity of SOM tend to increase tremendously leading to increased time consumption as well. Our work focuses on the analysis of different weight initialization strategies and various dimensionality reduction measures with the intent to make SOM flexible for handling high-dimensional datasets. We use two methods of comparison, one on projected space and another before projection. The datasets used are real world datasets taken from UCI repository.