الگوریتم های چند جانبه برای تشخیص تقلب

This paper describes an application of Case-Based Reasoning to the problem of reducing the number of final-line fraud investigations in the credit approval process. The performance of a suite of algorithms, which are applied in combination to determine a diagnosis from a set of retrieved cases, is reported. An adaptive diagnosis algorithm combining several neighbourhood-based and probabilistic algorithms was found to have the best performance, and these results indicate that an adaptive solution can provide fraud filtering and case ordering functions for reducing the number of final-line fraud investigations necessary.

Artificial intelligence techniques have been successfully applied to credit card fraud detection and credit scoring, and the field of AI as applied to the financial domain is both well-developed and well documented. As an emerging methodology, case-based reasoning (CBR) is now making a significant contribution to the task of fraud detection. CBR systems are able to learn from sample patterns of credit card use to classify new cases, and this approach also has the promise of being able to adapt to new patterns of fraud as they emerge. At the forefront of research in this field is the application of adaptive and hybrid learning systems to problems which previously were considered too dynamic, chaotic, or complex to accurately model and predict. As applied to the financial domain, CBR systems have a number of advantages over other AI techniques as they: • provide meaningful confidence and system accuracy measures; • require little or no direct expert knowledge acquisition; • are easily updated and maintained; • articulate the reasoning behind the decision making clearly; • are flexible and robust to missing or noisy data; • may take into account the cost effectiveness ratio of investigating false positives and advise accordingly; and • are easily integrated into varying database standards. And the addition of adaptive CBR components may allow the system to: • optimise the accuracy of classification by dynamically adjusting and updating weighting structures; • use multiple algorithms to enhance final diagnostic accuracy; and • better differentiate between types of irregularities and develop a diagnostically significant sense of abnormality which aids in the first-time detection of new irregularity types. In this paper we describe the background of a complex fraud-finding task, and then describe the development of an adaptive proof-of-concept CBR system, which is able to achieve very encouraging results on large, noisy real-world test sets. In specific, this paper addresses the problem of making a diagnostic decision given a set of near-matching cases. Finally, the results of the investigation are summarised and considered in the light of other work in the field.

Investigations into the financial data provided has proven that, though highly chaotic, it has properties that allow multi-algorithmic and adaptive CBR techniques to be used for fraud classification and filtering. The data set could not be partitioned into fraud and non-fraud regions. Instead, the occurrence and distribution of fraud cases in the neighbourhood of an unknown application was observed to be diagnostically significant, and these relationships were effectively exploited by a multi-algorithmic proof-of-concept system. Neighbourhood-based and probabilistic algorithms have been shown to be appropriate techniques for classification, and may be further enhanced using additional diagnostic algorithms for decision making in borderline cases, and for calculating confidence and relative risk measures. While more accurate performance metrics and more thorough testing is required to appropriately quantify peak precision, the initial testing results of 80% non-fraud and 52% fraud recognition strongly suggest that a multi-algorithmic CBR will be capable of high accuracy rates. A comparison with related work shows that CBR techniques can achieve similar performance in comparable problem areas. We believe that these results are very promising and supportive of a multi-algorithmic approach to classifying and assessing large, noisy data sets, and future work will focus upon testing the algorithms and resolution strategies on similarly complex data sets from other real-world domains.