A weight initialization approach for training Self Organizing Maps for clustering applications

The strength of Self Organizing Map (SOM) learning algorithm completely depends on the weights adjustments done in its network. Prior to the weight adjustments done, important step is to initialize the values to the weight. The choice of these initial values for weight vectors affects the performance of SOM training when applied to clustering. This paper proposes a different approach for initializing SOM. This approach depends on Frequency Sensitive Competitive Learning (FSCL) algorithm to pre-process the weights in order to improve the results obtained from trained input patterns in terms of better neuron utilization and less quantization and topographic error. Two datasets are used to analyze the performance of SOM algorithm. First dataset is evenly distributed 2D Gaussian data and second dataset is taken from the well reputed Engineering Educational Organization. Applying existing approaches of weight initializations, results obtained with first dataset showed that decreasing learning rate to a specific value gives better performance further but with second dataset results did not improve on decreasing learning rate. But with this new approach, results showed significant improvement as compared to the existing approaches of weight initialization.