Authors : Asha Gowda Karegowda, Vidya T. Shama, M.A. Jayaram and A.S. Manjunath
Abstract: Medical data mining is the process of extracting hidden patterns from medical data. Among the various clustering algorithms, k-means is the one of most widely used clustering technique. The performance of k-means clustering depends on the initial cluster centers and might converge to local optimum. k-means does not guarantee unique clustering because it generates different results with randomly chosen initial clusters for different runs of k-means. In addition the performance of any data mining depends on feature subset selection. This study attempts to improve performance of k-means clustering using two stages. As part of first stage, this study investigates the use of wrapper approach for feature selection for clustering where Genetic Algorithm (GA) is used as a random search technique for subset generation, wrapped with k-means clustering. In second stage, GA and Entropy based Fuzzy Clustering (EFC) are used to find the initial centroid for k-means clustering. Experiments have been conducted using standard medical dataset namely Pima Indians Diabetes Dataset (PIDD) and Heart statlog. Results show markable reduction of 8.42 and 18.89% in the classification error of k-means clustering for PIDD and Heart statlog dataset using features identified by proposed wrapper approach and initial centroids identified by GA when compared to k-means performance with all the features and centroids initialized by random method for PIDD and Heart statlog dataset.
Asha Gowda Karegowda, Vidya T. Shama, M.A. Jayaram and A.S. Manjunath, 2012. Genetic Algorithm Based Dimensionality Reduction for Improving Performance of K-Means Clustering: A Case Study for Categorization of Medical Dataset. International Journal of Soft Computing, 7: 249-255.