Unsupervised Ensemble Learning for Mining Top-n Outliers

Outlier detection is an important and attractive problem in knowledge discovery in large datasets. Instead of detecting an object as an outlier, we study detecting the n most outstanding outliers, i.e. the top-n outlier detection. Further, we consider the problem of combining the top-n outlier lists from various individual detection methods. A general framework of ensemble learning in the top-n outlier detection is proposed based on the rank aggregation techniques. A scorebased aggregation approach with the normalization method of outlier scores and an order-based aggregation approach based on the distance-based Mallows model are proposed to accommodate various scales and characteristics of outlier scores from different detection methods. Extensive experiments on several real datasets demonstrate that the proposed approaches always deliver a stable and effective performance independent of different datasets in a good scalability in comparison with the state-of-the-art literature.