Prediction-based outlier detection with explanations

General outlier detection strategies, be a distribution-based, clustering-based, or distance-based method, all resort to the comparison among instances to define abnormality. In this paper we introduce an additional dimension into the outlier definition. That is, we not only consider externally how one instance differs from others but internally the dependency and abnormality among its own attributes, denoted as the prediction-based outlier detection. Prediction-based outliers possess certain attributes which are difficult to be predicted based on the neighborhood information. Furthermore, we propose three neighborhood functions to generate predictions. Finally, acknowledging the lack of the gold standard to evaluate an outlier detection system, we propose four general evaluation strategies. Experiments conducted on several real-world datasets demonstrate the validity, novelty, power-law distribution, and robustness of our method.