Rough K-means Outlier Factor Based on Entropy Computation

Abstract

Many studies of outlier detection have been developed based on the cluster-based outlier detection approach, since it does not need any prior knowledge of the dataset. However, the previous studies only regard the outlier factor computation with respect to a single point or a small cluster, which reflects its deviates from a common cluster. Furthermore, all objects within outlier cluster are assumed to be similar. The outlier objects intuitively can be grouped into the outlier clusters and the outlier factors of each object within the outlier cluster should be different gradually. It is not natural if the outlierness of each object within outlier cluster is similar. This study proposes the new outlier detection method based on the hybrid of the Rough K-Means clustering algorithm and the entropy computation. We introduce the outlier degree measure namely the entropy outlier factor for the cluster based outlier detection. The proposed algorithm sequentially finds the outlier cluster and calculates the outlier factor degree of the objects within outlier cluster. Each object within outlier cluster is evaluated using entropy cluster-based to a whole cluster. The performance of the algorithm has been tested on four UCI benchmark data sets and show outperform especially in detection rate.