A Learning-Based Ensemble Algorithm with Optimal Selection for Outlier Detection

In this paper, we propose a Learning-based Ensemble Method with Optimal selection strategy (LbEM-OSS), which presents a new outlier detection algorithm that captures only outstanding ones of constituent models. Using KNN to define local regions and Pearson correlation to evaluate the detectors makes the ensemble robust. Our method can adapt and generalize better across different high-dimensional datasets by generating pseudo-ground truths with average and maximum aggregation strategies. On a wide range of benchmark datasets, LbEM-OSS outperformed both statistics-based and neural ensemble methods, which achieved stateof-the-art ROC-AUC as high as 97.78% in the best-case and 4-8% AUC improvements over existing methods on average. These results portray its potential for noise, different dimensionality, and heterogeneous data nature. Moreover, it is highly scalable and accurate, which makes it an essential application in practical fields like fraud detection, network security, and healthcare. This research highlights the need for dynamic selection approaches within ensemble methods, providing the groundwork for future developments in sound outlier detection.