PBC4occ: A novel contrast pattern-based classifier for one-class classification

In addition to accuracy, another key desirable characteristic of a classifier is interpretability. While there have been attempts to design contrast pattern-based models that support competitive and understandable classifiers, the utility of contrast patterns on the one-class classification problem is an under-explored area. In this paper, we propose a novel pattern-based classifier for one-class classification problems, PBC4occ. Moreover, we introduce the first contrast pattern mining algorithm utilizing decision trees for one-class classification. We analyze a number of contrast patterns extracted by our proposal and the one-class decision boundary built from an explanatory point. Additionally, we compare the performance of our proposal and those of thirteen (13) other state-of-the-art one-class classifiers on 95 imbalanced databases. Our findings show that PBC4occ achieves the best average value for both AUC and EER metrics. In addition, our proposal achieves the best and the second-best average Friedman’s ranking when evaluated under EER and AUC, respectively.