Impact of K-Means on the Performance of Classifiers for Labeled Data

Abstract

In this study a novel framework for data mining in clinical decision making have been proposed. Our framework addresses the problems of assessing and utilizing data mining models in medical domain. The framework consists of three stages. The first stage involves preprocessing of the data to improve its quality. The second stage employs k-means clustering algorithm to cluster the data into k clusters (in our case, k=2 i.e. cluster0 / no, cluster1 / yes) for validation the class labels associated with the data. After clustering, the class labels associated with the data is compared with the labels generated by clustering algorithm if both the labels are same it is assumed that the data is correctly classified. The instances for which the labels are not same are considered to be misclassified and are removed before further processing. In the third stage support vector machine classification is applied. The classification model is validated by using k-fold cross validation method. The performance of SVM (Support Vector Machine) classifier is also compared with Naive Bayes classifier. In our case SVM classifier outperforms the Naive Bayes classifier. To validate the proposed framework, experiments have been carried out on benchmark datasets such as Indian Pima diabetes dataset and Wisconsin breast cancer dataset (WBCD).These datasets were obtained from the University of California at Irvine (UCI) machine learning repository. Our proposed study obtained classification accuracy on both datasets, which is better with respect to the other classification algorithms applied on the same datasets as cited in the literature. The performance of the proposed framework was also evaluated using the sensitivity and specificity measures.