روش درخت تصمیم گیری حساس به هزینه برای تشخیص کلاهبرداری

With the developments in the information technology, fraud is spreading all over the world, resulting in huge financial losses. Though fraud prevention mechanisms such as CHIP&PIN are developed for credit card systems, these mechanisms do not prevent the most common fraud types such as fraudulent credit card usages over virtual POS (Point Of Sale) terminals or mail orders so called online credit card fraud. As a result, fraud detection becomes the essential tool and probably the best way to stop such fraud types. In this study, a new cost-sensitive decision tree approach which minimizes the sum of misclassification costs while selecting the splitting attribute at each non-terminal node is developed and the performance of this approach is compared with the well-known traditional classification models on a real world credit card data set. In this approach, misclassification costs are taken as varying. The results show that this cost-sensitive decision tree algorithm outperforms the existing well-known methods on the given problem set with respect to the well-known performance metrics such as accuracy and true positive rate, but also a newly defined cost-sensitive metric specific to credit card fraud detection domain. Accordingly, financial losses due to fraudulent transactions can be decreased more by the implementation of this approach in fraud detection systems.

Fraud can be defined as wrongful or criminal deception aimed to result in financial or personal gain. The two main mechanisms to avoid frauds and losses due to fraudulent activities are fraud prevention and fraud detection systems. Fraud prevention is the proactive mechanism with the goal of disabling the occurrence of fraud. Fraud detection systems come into play when the fraudsters surpass the fraud prevention systems and start a fraudulent transaction. A review of fraud domains and detection techniques can be found in Bolton and Hand, 2002, Kou et al., 2004, Phua et al., 2005 and Sahin and Duman, 2010. One of the most well-known fraud domains is the credit card systems. Credit card frauds can be made in many ways such as simple theft, application fraud, counterfeit cards, never received issue (NRI) and online fraud (where the card holder is not present). In online fraud, the transaction is made remotely and only the card’s details are needed. Because of the international availability of the web and ease with which users can hide their location and identity over internet transactions, there is a rapid growth of committing fraudulent actions over this medium. There are many previous studies done on credit card fraud detection. The general background of the credit card systems and non-technical knowledge about this fraud domain can be learned from Hanagandi, Dhar, and Buescher (1996) and Hand and Blunt (2001), respectively. The most commonly used fraud detection methods in this domain are rule-induction techniques, decision trees, Artificial Neural Networks (ANN), Support Vector Machines (SVM), logistic regression, and meta-heuristics such as genetic algorithms. These techniques can be used alone or in collaboration using ensemble or meta-learning techniques to build classifiers. Most of the credit card fraud detection systems are using supervised algorithms such as neural networks (Brause et al., 1999, Dorronsoro et al., 1997, Juszczak et al., 2008, Quah and Sriganesh, 2008, Schindeler, 2006, Shen et al., 2007, Stolfo et al., 1997, Stolfo et al., 1999 and Syeda et al., 2002; Prodromidis, Chan, & Stolfo, 2000); decision tree techniques like ID3, C4.5 and C&RT (Chen et al., 2004, Chen et al., 2005, Mena, 2003 and Wheeler and Aitken, 2000); and SVM (Gartner Reports, 2010 and Leonard, 1993). Credit card fraud detection is an extremely difficult, but also popular problem to solve. There comes only a limited amount of data with the transaction being committed. Also, there can be past transactions made by fraudsters which also fit a pattern of normal (legitimate) behavior (Aleskerov, Freisleben, & Rao, 1997). Furthermore, the problem has many constraints. First of all, the profiles of normal and fraudulent behaviors change constantly. Secondly, the development of new fraud detection methods is made more difficult by the fact that the exchange of ideas in fraud detection, especially in credit card fraud detection is severely limited due to security and privacy concerns. Thirdly, data sets are not made available and the results are often censored, making them difficult to assess. Even, some of the studies are done using synthetically generated data (Brause et al., 1999 and Dorronsoro et al., 1997). Fourthly, credit card fraud data sets are highly skewed sets. Lastly, the data sets are also constantly evolving making the profiles of normal and fraudulent behaviors always changing (Bolton and Hand, 2002, Kou et al., 2004, Phua et al., 2005 and Sahin and Duman, 2010). So, credit card fraud detection is still a popular challenging and hard research topic. Visa reports about credit card frauds in European countries state that about 50% of the whole credit card fraud losses in 2008 are due to online frauds (Ghosh & Reilly, 1994). Many papers reported huge amounts of losses in different countries (Bolton and Hand, 2002, Dahl, 2006 and Schindeler, 2006). Thus new approaches improving the classifier performance in this domain have both financial implications and research contributions. Defining a new cost-sensitive approach is one of the best ways for such an improvement due to the characteristics of the domain. Although traditional machine learning techniques are generally successful in many classification problems, having a high accuracy or minimizing the misclassification errors is not always the goal for the classifier developed. In the applications of real-world machine-learning problem domains, there are various types of costs involved and Turney defined nine main types of costs (Turney, 2000). However, most of the machine-learning literature does not take any of these costs into account, only a few of the remainings take the misclassification cost into consideration. Turney also stated that the cost of misclassification errors occupies a unique position in their taxonomy (Turney, 2000). Nevertheless, according to the Technological Roadmap of the ML-netII project (European Network of Excellence in Machine Learning), cost-sensitive learning is stated to be one of the most popular topics in the future of machine learning research (Saitta, 2000 and Zhou and Liu, 2006). Thus, improving classifier performance of a fraud detection system by building cost-sensitive classifiers is the best way to enable recovery of large amounts of financial losses. Besides, the customer loyalty and trust will also be increased. Also, cost-sensitive classifiers have been shown to be effective in addressing the class imbalance problem (Thai-Nghe et al., 2010 and Zhou and Liu, 2006). Most of the past studies work on constant misclassification cost matrices or cost matrices composed of a number of constant heterogeneous misclassification costs; however, each false negative (FN) has a unique misclassification cost inherent to it. Accordingly, each FN should be prioritized in some way to show the misclassification cost difference. For example, a fraudulent transaction with a larger transaction amount or a larger usable card limit should be more important to detect than one with a smaller amount or usable card limit. A constant cost matrix or a combination of constant cost matrices cannot depict this picture. So, this study is one of the pioneers to take such cases into account while working with classification problem under variable misclassification costs. This is one of the gaps in the literature of credit card fraud detection aimed to be filled by this study. In this study, a new cost-sensitive decision tree induction algorithm that minimizes the sum of misclassification costs while selecting the splitting attribute at each non-terminal node of the tree is developed and the classification performance is compared with those of the traditional classification methods, both cost-insensitive and cost-sensitive with fixed misclassification cost ratios, such as traditional decision tree algorithms, ANN and SVM. The results show that this cost-sensitive decision tree algorithm outperforms the existing well-known methods on our real-world data set in terms of the fraudulent transactions identified and the amount of possible losses prevented. In credit card fraud detection, the misclassification costs and the priorities of the frauds to be identified differ depending on the individual records. As a result, the common performance metrics such as accuracy, True Positive Rate (TPR) or even Area Under Curve are not suitable to evaluate the performance of the models because they accept each fraud as having the same priority regardless of the amount of that fraudulent transaction or the available usable limit of the card used in the transaction at that time. A new performance metric which prioritizes each fraudulent transaction in a meaningful way and checks the performance of the model in minimizing the total financial loss should be used. Fraudsters generally deplete the usable limit of a credit card once they get the opportunity of committing fraudulent transactions using the card. Accordingly, the financial loss of a fraudulent transaction can be assumed as the available limit of the card before the transaction instead of the amount of the transaction. So, the performance comparisons of the models over the test set are done over the newly defined cost-sensitive performance metric Saved Loss Rate (SLR) which is the saved percentage of the potential financial loss that is the sum of the available usable limits of the cards from which fraudulent transactions are committed. To show the correctness of our argument, True Positive Rate (TPR) values for the performance of the models are also given in performance comparisons of the models. The rest of this paper is organized as follows: Section 2 gives a review of the cost-sensitive approaches in machine learning and Section 3 gives some insights to the structure of credit card data. Section 4 gives the details of the newly developed cost-sensitive decision tree algorithm. Section 5 gives the results and a short discussion about the results and Section 6 concludes the study.

In this work, we have developed and implemented a number of cost-sensitive decision tree approaches to be used in credit card fraud detection and show that it outperforms the models built using the traditional data mining methods such as decision trees, ANN and SVM. We propose a new approach for making the splits by taking variable misclassification costs depending on each individual record into account. The performances of the models are compared on a real-world data set showing that the models can be readily implemented in real-world systems. We evaluated the proposed algorithm on a real world data set and demonstrated the conclusions which can be used as guidelines for working with a classification problem where variable misclassification cost depending on each individual example should be taken into account. We revealed that the well-known performance metrics such as accuracy and TPR are not suitable for this kind of problems, and developed a new performance metric for the credit card fraud detection problem which is the percentage of available usable limits saved. The performances of classifiers built using CS – Direct Cost method on real-world test set indicate that we cannot use misclassification cost without incorporating the class distribution or an impurity measure in cost calculations. However, using our cost-sensitive approaches which incorporate such an information in the cost calculations make a significant improvement in the classification performance both with respect to TPR and the newly defined domain specific metric SLR over the best existing methods. We believe that this new metric will be a widely accepted and adopted in the future research studies in credit card fraud detection. These performance improvements indicate many research contributions and managerial implications. First of all, research on classification in domains with imbalanced data such as fraud detection, specifically credit card fraud detection or medical diagnosis where misclassification costs highly differ among classes should focus on cost-sensitive classification to build classifiers which can prioritize minority class instances with big misclassification costs. Although there are many studies on cost-sensitive modeling in medical diagnosis, our study, as long as we know, is one of the pioneers to combine cost-sensitive modeling using decision trees and credit card fraud detection. Furthermore, it will serve as a guide for further studies in cost-sensitive classification. Secondly, by implementing such algorithms in real world fraud detection systems, financial institutions can save huge amounts of money when the financial losses due to fraud are considered. An improvement in classification performance less than 1% may result in millions of dollar savings. The performance difference among cost-sensitive models and traditional models in our test data corresponds to a huge amount. Such a performance improvement in larger banks with many customers will result in more savings. Accordingly, financial institutions are in a search for such improvements. Also, the loyalty and trust of customers in financial institutions will be enriched by the way.