An investigation of real-valued accuracy-based learning classifier systems for electronic fraud detection

Abstract

Fraud is a serious problem that costs the worldwide economy billions of dollars annually. However, fraud detection is difficult as perpetrators actively attempt to masquerade their actions, among typically overwhelming large volumes of, legitimate activity. In this paper, we investigate the fraud detection problem and examine how learning classifier systems can be applied to it. We describe the common properties of fraud, introducing an abstract problem which can be tuned to exhibit those characteristics. We report experiments on this abstract problem with a popular real-time learning classifier system algorithm; results from our experiments demonstrating that this approach can overcome the difficulties inherent to the fraud detection problem. Finally we apply the algorithm to a real-world problem and show that it can achieve good performance in this domain.

PCA for improving the performance of XCSR in classification of high-dimensional problems

“…Marín-Blázquez and Martínez Pérez [10] use a modified version of XCS called linguistic hedged fuzzy-XCS to tackle the KDD'99 problem; the advantage of their approach is that their system returns a set of human interpretable knowledge (rules). In [2], Behdad et al apply XCSR on the KDD'99 data-set as an example of a realworld fraud detection problem and report competitive performance.…”

“…In our previous work [2], we used XCSR as a network intrusion detection system. We considered both online learning (where each new instance from the test set may influence the set of classifiers) and offline learning (where the set of classifiers is derived from a training set and then statically applied to a test set).…”

“…We considered both online learning (where each new instance from the test set may influence the set of classifiers) and offline learning (where the set of classifiers is derived from a training set and then statically applied to a test set). Our resultant system was evaluated against the KDD'99 data-set [6], with results indicating performance was comparable to that achieved by the winner of the KDD'99 competition without online learning, and superior to the performance of the KDD'99 winner when online learning was used [2].…”

“…In our previous work [2], the population size that gave the best results for XCSR on KDD'99 was 50, 000. So, we use the same population size for XCSR.…”

“…So, we use the same population size for XCSR. The other parameter settings for XCSR are the same as the one in [2]. All the experiments are repeated 10 times, and the results averaged.…”