Pattern classification algorithms usually assume, that the distribution of examples in classes is roughly balanced. However, in many cases one of the classes is dominant in comparison with others. Here, the classifier will become biased towards the majority class. This scenario is known as imbalanced classification. As the minority class is usually the one more valuable, we need to counter the imbalance effect by using one of several dedicated techniques. Cost-sensitive methods assume a penalty factor for misclassifying the minority objects. This way, by assuming a higher cost to minority objects we boost their importance for the classification process. In this paper, we propose a model of cost-sensitive neural network with moving threshold. It relies on scaling the output of the classifier with a given cost function. This way, we adjust our support functions towards the minority class. We propose a novel method for automatically determining the cost, based on the Receiver Operating Characteristic (ROC) curve analysis. It allows us to select the most efficient cost factor for a given dataset. Experimental comparison with state-of-the-art methods for imbalanced classification and backed-up by a statistical analysis prove the effectiveness of our proposal.
CITATION STYLE
Krawczyk, B., & Woźniak, M. (2015). Cost-sensitive neural network with ROC-based moving threshold for imbalanced classification. In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) (Vol. 9375 LNCS, pp. 45–52). Springer Verlag. https://doi.org/10.1007/978-3-319-24834-9_6
Mendeley helps you to discover research relevant for your work.