SAD: A novel method for ensemble outlier detection with dynamic prediction label

Abstract

Majority voting outlier detection is a traditional method that has been widely used in many fields. It uses the strategy of majority vote to make a prediction, which makes it perform poorly in acc index sometimes. In this paper, a method called second anomaly detection (SAD) is proposed, to detect the connection of outlier scores between each other and decide the advantage strength of a sample when defining the outlierness, which is expressed as $a$ factor, then the prediction label of a sample is ascertained according to the a value. Finally, SAD is compared with several majority voting anomaly detection algorithms in accuracy performance, such as iForest, HBOS, AutoEncoder, it is shown that the proposed algorithm SAD is effective.