Using multiple classifier behavior to develop a dynamic outlier ensemble

Outlier ensembles that use more base detectors recently become an attractive approach to solving problems of single detectors. However, existing outlier ensembles often assume that base detectors make independent errors, which is difficult to satisfy in practical applications. To this end, this paper proposes a dynamic outlier ensemble to loose this error independence assumption. In our method, it is desired that the most competent base detector(s) can be singled out by the dynamic selection mechanism for each test pattern. The usage of the concept of multiple classifier behavior (MCB) has two purposes. One is to generate artificial outlier examples used for competence estimates. This strategy is different from other methods since we do not make any assumption regarding the data distribution. On the other hand, MCB is used to refine validation sets initialized by the K-nearest neighbors (KNN) rule. It is desired that objects in the refined validation sets are more representative than those found by KNN. With the refined validation sets, competences of all base detectors will be estimated by a probabilistic method, before which we have transformed outputs of base detectors into a probabilistic form. Finally, a switching mechanism that determines whether one detector should be nominated to make the decision or a fusion method should be applied instead is proposed in order to achieve a robust detection result. We carry out experiments on 20 benchmark data sets to verify the effectiveness of our detection method.