Two Outlier-Sensitive Measures for Semi-supervised Dynamic Ensemble Anomaly Detection Models

Semi-supervised anomaly detection has received wide interest because of not requiring counterexamples during training. Existing competence measures for semi-supervised dynamic ensemble anomaly detection models do not consider the imbalance characteristic of training samples, which will result in serious overfitting on normal samples. This paper proposes two outlier-sensitive measures to estimate the competence of base classifiers for dynamic ensemble models. When a normal sample is correctly classified, both measures give a higher positive score to base classifiers with confidence closer to 0.5, which is different from the conventional idea that base classifiers with higher confidence should obtain higher scores. When a sample is misclassified, the Output-based Outlier-Sensitive measure calculates a negative score based on the confidence outputted by the base classifier, while the Cost-Sensitive-based Outlier-Sensitive measure gives a negative score based on the category of this sample. Multiple experiments are carried out on 30 datasets from public repositories under the unified framework proposed in this paper, and results show that dynamic ensemble models with our competence measures can outperform a number of typical ensemble models in terms of G-mean and F1, regardless of the pseudo outlier labeling methods and base classifier selection methods used in the model.