Outlier Detection Method based on Improved Two-step Clustering Algorithm and Synthetic Hypothesis Testing

摘要

For the detection of outliers in unsupervised mixed attribute datasets, this paper proposes an outlier detection method based on improved two-step clustering algorithm and synthetic hypothesis testing. Firstly, a two-step clustering algorithm based on clustering feature tree is improved to be suitable for mixed attribute datasets. In the clustering stage, RA-Clust algorithm is used to pre-select the clustering center to improve the algorithm performance, and the original complex data distribution is simplified into the superposition of data distribution under sub-clusterings. Then using the method of synthetic hypothesis testing, the outliers are judged according to the change rule of data in the sub-clusterings of sample data. Experimental results show that the algorithm has better clustering performance, and the synthetic hypothesis testing can accurately detect most of the outliers and maintain over 90% of the outlier detect efficiency under different data sizes.