Development of hybrid classification methodology for mining skewed data sets: A case study of Indian customs data

Abstract

At present, detecting customs declaration frauds with limited examination of imported goods by available scarce resources is posing considerable challenge to the customs authorities world over. Data mining techniques could be utilized to sift through the past data and develop predictive model for examination of limited goods with higher probability of fraud. However, this requires handling large, skewed data sets with variable error of each misclassification. Literature suggests various data level and algorithm level interventions for addressing these issues. Successive application of combination of both the types of interventions on the classification tree technique is devised in this paper to improve the predictive accuracy of the model. Furthermore, the predictions of this classification tree model are then fed into an artificial neural classification model, which gives the flexibility to modulate the predictive accuracy of a particular class label to suit the end objective. This methodology can be effectively applied to other similar situations such as detecting insurance fraud, credit card fraud, telecom churning and frauds etc. © 2006 IEEE.