While working on different fraud detection projects, I developed my own approach to the solution for this problem. In my PASS Summit 2013 session I am introducing this approach. I also wrote a whitepaper on the same topic, which was generously reviewed by my friend Matija Lah. In order to spread this knowledge faster, I am starting a series of blog posts which will at the end make the whole whitepaper.
With the massive usage of credit cards and web applications for banking and payment processing, the number of fraudulent transactions is growing rapidly and on a global scale. Several fraud detection algorithms are available within a variety of different products. In this paper, we focus on using the Microsoft SQL Server suite for this purpose. In addition, we will explain our original approach to solving the problem by introducing a continuous learning procedure. Our preferred type of service is mentoring; it allows us to perform the work and consulting together with transferring the knowledge onto the customer, thus making it possible for a customer to continue to learn independently.
This paper is based on practical experience with different projects covering online banking and credit card usage.
A fraud is a criminal or deceptive activity with the intention of achieving financial or some other gain. Fraud can appear in multiple business areas. You can find a detailed overview of the business domains where fraud can take place in
Sahin Y., & Duman E. (2011), Detecting Credit Card Fraud by Decision Trees and Support Vector Machines, Proceedings of the International MultiConference of Engineers and Computer Scientists 2011 Vol 1. Hong Kong: IMECS.
Dealing with frauds includes fraud prevention and fraud detection. Fraud prevention is a proactive mechanism, which tries to disable frauds by using previous knowledge. Fraud detection is a reactive mechanism with the goal of detecting suspicious behavior when a fraudster surpasses the fraud prevention mechanism. A fraud detection mechanism checks every transaction and assigns a weight in terms of probability between 0 and 1 that represents a score for evaluating whether a transaction is fraudulent or not. A fraud detection mechanism cannot detect frauds with a probability of 100%; therefore, manual transaction checking must also be available. With fraud detection, this manual part can focus on the most suspicious transactions. This way, an unchanged number of supervisors can detect significantly more frauds than could be achieved with traditional methods of selecting which transactions to check, for example with random sampling.
There are two principal data mining techniques available both in general data mining as well as in specific fraud detection techniques: supervised or directed and unsupervised or undirected. Supervised techniques or data mining models use previous knowledge. Typically, existing transactions are marked with a flag denoting whether a particular transaction is fraudulent or not. Customers at some point in time do report frauds, and the transactional system should be capable of accepting such a flag. Supervised data mining algorithms try to explain the value of this flag by using different input variables. When the patterns and rules that lead to frauds are learned through the model training process, they can be used for prediction of the fraud flag on new incoming transactions. Unsupervised techniques analyze data without prior knowledge, without the fraud flag; they try to find transactions which do not resemble other transactions, i.e. outliers. In both cases, there should be more frauds in the data set selected for checking by using the data mining knowledge compared to selecting the data set with simpler methods; this is known as the lift of a model. Typically, we compare the lift with random sampling. The supervised methods typically give a much better lift than the unsupervised ones. However, we must use the unsupervised ones when we do not have any previous knowledge. Furthermore, unsupervised methods are useful for controlling whether the supervised models are still efficient.
Accuracy of the predictions drops over time. Patterns of credit card usage, for example, change over time. In addition, fraudsters continuously learn as well. Therefore, it is important to check the efficiency of the predictive models with the undirected ones. When the difference between the lift of the supervised models and the lift of the unsupervised models drops, it is time to refine the supervised models. However, the unsupervised models can become obsolete as well. It is also important to measure the overall efficiency of both, supervised and unsupervised models, over time. We can compare the number of predicted frauds with the total number of frauds that include predicted and reported occurrences. For measuring behavior across time, specific analytical databases called data warehouses (DW) and on-line analytical processing (OLAP) systems can be employed. By controlling the supervised models with unsupervised ones and by using an OLAP system or DW reports to control both, a continuous learning infrastructure can be established.
There are many difficulties in developing a fraud detection system. As has already been mentioned, fraudsters continuously learn, and the patterns change. The exchange of experiences and ideas can be very limited due to privacy concerns. In addition, both data sets and results might be censored, as the companies generally do not want to publically expose actual fraudulent behaviors. Therefore it can be quite difficult if not impossible to cross-evaluate the models using data from different companies and different business areas. This fact stresses the importance of continuous learning even more. Finally, the number of frauds in the total number of transactions is small, typically much less than 1% of transactions is fraudulent. Some predictive data mining algorithms do not give good results when the target state is represented with a very low frequency. Data preparation techniques like oversampling and undersampling can help overcome the shortcomings of many algorithms.
SQL Server suite includes all of the software required to create, deploy any maintain a fraud detection infrastructure. The Database Engine is the relational database management system (RDBMS), which supports all activity needed for data preparation and for data warehouses. SQL Server Analysis Services (SSAS) supports OLAP and data mining (in version 2012, you need to install SSAS in multidimensional and data mining mode; this was the only mode in previous versions of SSAS, while SSAS 2012 also supports the tabular mode, which does not include data mining). Additional products from the suite can be useful as well. SQL Server Integration Services (SSIS) is a tool for developing extract transform–load (ETL) applications. SSIS is typically used for loading a DW, and in addition, it can use SSAS data mining models for building intelligent data flows. SQL Server Reporting Services (SSRS) is useful for presenting the results in a variety of reports. Data Quality Services (DQS) mitigate the occasional data cleansing process by maintaining a knowledge base. Master Data Services is an application that helps companies maintaining a central, authoritative source of their master data, i.e. the most important data to any organization. For an overview of the SQL Server business intelligence (BI) part of the suite that includes Database Engine, SSAS and SSRS, please refer to
Veerman E., Lachev T., & Sarka D. (2009). MCTS Self-Paced Training Kit (Exam 70-448): Microsoft® SQL Server® 2008 Business Intelligence Development and Maintenance. MS Press.
For an overview of the enterprise information management (EIM) part that includes SSIS, DQS and MDS, please refer to
Sarka D., Lah M., & Jerkič G. (2012). Training Kit (Exam 70-463): Implementing a Data Warehouse with Microsoft® SQL Server® 2012. O'Reilly.
For details about SSAS data mining, please refer to
MacLennan J., Tang Z., & Crivat B. (2009). Data Mining with Microsoft SQL Server 2008. Wiley.
SQL Server Data Mining Add-ins for Office, a free download for Office versions 2007, 2010 and 2013, bring the power of data mining to Excel, enabling advanced analytics in Excel. Together with PowerPivot for Excel, which is also freely downloadable and can be used in Excel 2010, is already included in Excel 2013. It brings OLAP functionalities directly into Excel, making it possible for an advanced analyst to build a complete learning infrastructure using a familiar tool. This way, many more people, including employees in subsidiaries, can contribute to the learning process by examining local transactions and quickly identifying new patterns.