University of Houston

We propose a pre-processing step to classification that applies a Self Organising Map(SOM) algorithm to the training set to discover patterns in the attribute or input space. We demonstrate how this knowledge can be exploited to enhance the predictive accuracy of Support Vector Machines (SVM). Our approach serves as a framework to improve the performance of classifiers that exhibitpoor performance when the dataset is characterized by many clusters per class. The 2D graphical representation of a SOM grid architecture helps to identify the clusters in need for relabeling. During classification, global and local learning are applied respectively to enhance accuracy performance. Experimental results on synthetic and real-world domains show an advantage in predictive accuracy. when class reconstruction is used as a pre-processing step to classification.