A novel approach for outlier detection and clustering improvement

Outlier detection is used to detect abnormalities in various application domains including clustering based disease onset identification, gene expression analysis, computer network intrusion, financial fraud detection and human behaviour analysis. Existing methods to detect outliers are inadequate due to poor accuracy and lack of any general technique. Most techniques consider either small clusters as outliers or provide a score for being outlier to each data object. These approaches have limitations due to high computational complexity and misidentification of normal data object as outliers. In this paper, we provide a novel unsupervised approach to detect outliers using a modified k-means clustering algorithm. The detected outliers are removed from the dataset to improve clustering accuracy. We validate our approach by comparing against existing techniques and benchmark performance. Experimental results on benchmark datasets show that our proposed technique outperforms existing methods on several measures.