A Roc Curve Based K-Means Clustering for Outlier Detection Using Dragon Fly Optimization

Abstract

Outlier detection is an essential step in the data mining process. Its main purpose to remove the incompatible data from the original data. This process helps in the removal of data which are necessary for carrying out to speed up the applications like classification, data perturbation and compression. It plays an important role in the weather forecasting, performance analysis of sports person and network intrusion detection systems. The outlier for the single variable can be easily observed but for the n-variable it become a tedious process. To enhance the performance of outlier detection in n-variable or attributes several methods were proposed. Some of the existing methods are statistical approaches, proximity-based measures, classification approaches, and index-based approaches and optimization based approaches. The first four approaches were not able to classify the data when there is an imperfection in the labels. But, the optimization based approach is able to overcome this problem even there is an imperfect labelling. One of the existing optimization approach is K-means and K-median based approach. The existing method failed to process the larger records and smaller attributes. To overcome this problem a dragon fly based K-means clustering and multi-layer feed forward neural network is proposed. This objective is achieved with the help of ROC curve (negative ratio) as objective function. The performance is evaluated using detection rate. The proposed method is tested on datasets like Arrhythmia, Diabetics and Epileptic seizure which has larger attributes and larger records. The proposed method outperforms the existing approach with high detection rate of 0.95 for arrhythmia dataset and 0.96 for Epileptic dataset.