File(s) under permanent embargo
MEOD: A Robust Multi-stage Ensemble Model Based on Rank Aggregation and Stacking for Outlier Detection
conference contributionposted on 2023-02-23, 00:31 authored by Z Jiang, F Zhang, H Xu, L Tao, Zili ZhangZili Zhang
In ensemble-based unsupervised outlier detection, the lack of ground truth makes the combination of basic outlier detectors a challenging task. The existing outlier ensembles usually use certain fusion rules (majority voting, averaging) to aggregate base detectors, which results in relatively low model accuracy and robustness. To overcome this problem, in this research, we propose a robust Multi-stage Ensemble model based on rank aggregation and stacking for Outlier Detection (MEOD). The proposed model uses multiple unsupervised outlier detection algorithms to form a base detector pool. Such a pool can be utilized for extracting useful representations from the train set and integrating base detector results using a ranking aggregation-based approach. To further optimize the proposed model, a stacking-based dynamic classifier selection ensemble model is also proposed, and the best-behaved classifier is adaptively selected as the base learner in the stacking stage on different datasets. Some extensive experiments are also committed to prove that MEOD outperforms the other seven state-of-the-art benchmarks in most cases.