An Accuracy-and-Diversity-based Ensemble Method for Concept Drift and Its application in Fraud Detection

Concept Drift is one of the most challenging issues in various applications such as fraud detection, spam filtering and sensor networks, causing the performance degradation of learning algorithms which are dedicated to processing static data such as Naive Bayes, due to their poor adaptability to changes. Several techniques and algorithms have been proposed to address the concept drift problem, e.g. the ensemble methods that have already achieved great success. However, most of the ensemble methods consider either accuracy or diversity of a trained model when adapting to new concepts, and thus they can hardly deal with different types of drifts. This paper proposes a novel ensemble method that considers both accuracy and diversity. We linearly combine accuracy and diversity with different coefficients, and propose an entropy-based policy to compute these coefficients. Experiments on 10 synthetic datasets and 4 public datasets demonstrate that our method outperforms the state-of-the-art ones on dealing with different concept drifts. Our method is also applied to the electronic transaction fraud detection and achieves an excellent performance.