A new framework for combining supervised and semi-supervised methods in fraud detection

A new framework for combining supervised and semi-supervised methods in fraud detection

Abstract

Every year a large amount of money is lost due to fraud in financial institutions. Detecting frauds is a complicated task and limiting fraud detection systems to certain kinds of detection methods like supervised or unsupervised methods does not seem efficient. In this paper, a combined framework for fraud detection systems, consisting of both supervised and semi-supervised methods in three main components namely rule-based component, trend-analysis-based component and, a scenario-based component is proposed. The rule-based component is the supervised part of the framework and a decision tree, which is a cost-insensitive classification algorithm, is used for this component. In the trend-analysis-based component, which is the semi-supervised part of our proposed framework, the normal behavior of users are modeled and the extent of dissimilarities of newly arrived transactions are calculated. Finally, in the scenario-based component, which is another semi-supervised part of the proposed framework, the extent of similarities of the sequence of transactions with the known fraud scenarios are calculated. The final result is obtained through combining the results of all these three components using a bagging method. Combining the outputs of all these components together using the proposed bagging model, rather than detecting more frauds, the results are more stable and the number of false alarms is reduced remarkably.

References