Adaptive Approaches to Electronic Fraud Detection in Highly Dynamic Environments
As more and more companies and government agencies move towards electronic processing, they become more vulnerable to large scale and systematic fraud. This trend has led to a significant research effort towards providing algorithms and methods for fraud detection. These efforts have not been fully successful due to the unique characteristics of fraud detection; the most important of which is the adaptive environment. In other words, as the fraud detection techniques improve, the fraudsters change their behaviour. Other challenging characteristics of fraud detection are: imbalanced data, unequal misclassification costs, concept drift, overwhelming large volume of data, the required high accuracy, the required fast processing time and the lack of sufficient amount of training data.
LCS (Learning Classifier Systems) is an adaptive machine learning technique which combines reinforcement learning and evolutionary computing. It has been successfully applied in the areas such as modelling, robotics and data mining. However, despite having the very important feature of adaptability, it has not been extensively applied to fraud detection. The reason is that it is susceptible to imbalanced data and large volumes of data. The aim of this research is to tune and use LCS to detect electronic fraudulent. The improvements in the LCS algorithms will then be generalized and will readily be applicable to other domains.
As electronic technologies have become omnipresent, so has electronic fraud. In response to an increasing number of fraudulent incidents, researchers and security companies have come up with methods to detect such activities. But, fraudsters tend to be intelligent and constantly change their methods in order to evade detection. Therefore, there is a need for self-adaptive fraud detection approaches.
LCS (Learning Classifier Systems) is an adaptive rule-based learning framework which has the potential to be a suitable choice for fraud detection; and has not been considered for this purpose extensively before. One reason is that it is relatively susceptible to some properties of fraudulent activities, such as imbalanced classes. This research will tune LCS for fraud detection, and eventually produce a variant which performs well in such environments.